[iterate-3AL] Superblock dispatch: chain basic blocks per slot-visit (~1.6x boot-to-splash)

Replace the one-basic-block-per-slot-per-round lockstep dispatch with a
SUPERBLOCK runner: each slot-visit chains straight-line blocks through
their terminating branches up to a deterministic instruction budget,
amortizing the per-round (timebase/coord/round_schedule) and per-slot
(worker_prologue) dispatch tax over ~128 instructions instead of ~6.

Yield-points (end the chain, return to the round) are pure functions of
guest state, preserving the lockstep cross-thread interleaving correctness:
  - non-Continue step result (Yield/SystemCall/Trap/Unimpl/Halted);
    db16cyc Yield is the spin-wait producer hand-off.
  - sync-sensitive block: lwarx/ldarx/stwcx./stdcx. or sync/eieio/isync
    (new PpcOpcode::is_sync_sensitive, flagged on DecodedBlock at build).
  - MMIO touch: new GuestMemory::mmio_access_count() watermark, sampled
    per block, keeps GPU/register ordering at one-block granularity.
  - next PC leaves ordinary guest code (import thunk / halt sentinel /
    unmapped) -> hand to the full worker_prologue next round.
  - instruction budget reached.

Instruction-count/clock accounting stays exact: per-block cycle_count
deltas are summed and handed to worker_epilogue once (instruction_count +
decrement_quantum advance by the precise retired count). XENIA_SUPERBLOCK_BUDGET=1
reproduces the old one-block schedule byte-for-byte.

Budget tuned to 128 (env-overridable): boot progression stays healthy up
to 256, sharp cliff at ~384 (a boot producer/consumer handoff starves);
128 is 3x below the cliff. Also scale the inline-GPU per-round fairness
cap with the budget (flat 64 throttled GPU command processing 17x under
superblocks and collapsed the present loop).

PERF (check -n 100M --gpu-inline): 25.3 -> 42.7 MIPS (1.69x); 1B: 26.0 ->
41.4 MIPS (1.59x). Callgrind n=5M: host instructions 2.178B -> 1.507B
(-31%); worker_prologue -90%, coord_pre_round -91%, begin_slot_visit /
round_schedule_into / coord_post_round / update_timestamp_bundle each
~-90%; interpreter execute byte-identical (real work unchanged).

GATES: C1 boot progression 150M draws 7391/swaps 2164 (baseline 7415/2172),
1B draws 88547/swaps 29228 linear no stall, K8888 decode + RTs=2 intact.
C2 determinism: n50m stable digest byte-identical across fresh runs;
golden re-baselined intentionally (pacing-only deltas: imports 333453->243387,
draws 1274->1279). C3 milestone-1 render: texture_decodes/draws/swaps/
present cadence track baseline (3AJ fade-in pacing preserved). C4: 690
tests green (+2 sync_sensitive).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

crates/xenia-app/src/main.rs

@@ -1540,8 +1540,19 @@ fn cmd_exec_inner(
                     mem.write_u32(addr, block);
                 }
                 ("xboxkrnl.exe", 0x01BE) => {
-                    // VdGlobalDevice — passed through to Vd* shims. Write 0.
-                    mem.write_u32(addr, 0);
+                    // VdGlobalDevice — a *pointer to* a global D3D-device cell.
+                    // Mirror xenia-canary RegisterVideoExports (xboxkrnl_video.cc:
+                    // 557-564): allocate a 4-byte cell, point the import slot at
+                    // it, and zero the cell. The guest's graphics init then stores
+                    // its device object INTO the cell (e.g. sub_824C6DC0 @
+                    // 0x824C6F18 `stw r31, 0([0x82000750])`), and the swap-complete
+                    // callback sub_824CE2B8 reads it back via the two-level
+                    // `[[VdGlobalDevice]+0]+15160` to bump the swap counter (clock
+                    // B). Writing 0 directly here (the old behaviour) made that
+                    // store land at address 0 and the swap counter never advance —
+                    // freezing the title-loop's per-frame manager update.
+                    let cell = alloc_zero(0x4, &mut mem, &mut kernel);
+                    mem.write_u32(addr, cell);
                 }
                 ("xboxkrnl.exe", 0x01C0) => {
                     // VdGpuClockInMHz
@@ -2140,7 +2151,13 @@ fn coord_pre_round(
     let fired = if kernel.parallel_active {
         kernel.interrupts.tick_vsync_wallclock()
     } else {
-        kernel.interrupts.tick_vsync_instr(stats.instruction_count)
+        // iterate-3AJ: present-anchored — pass the guest's live present
+        // (`VdSwap`) count so vsync tracks the real present rate once the
+        // guest is presenting (≈1 vblank/present), instead of firing a
+        // fixed instruction quantum that over-fires ~66× during one heavy
+        // splash asset-load frame and collapsed the logo fade-in.
+        let presents = kernel.gpu.swaps_seen();
+        kernel.interrupts.tick_vsync_instr(stats.instruction_count, presents)
     };
     if fired {
         use std::sync::atomic::Ordering;
@@ -2309,8 +2326,19 @@ fn coord_post_round(
     let mut gpu_runs = (executed_this_round
         / xenia_cpu::scheduler::HW_THREAD_COUNT as u64)
         .max(1);
-    if gpu_runs > 64 {
-        gpu_runs = 64;
+    // Fairness cap on GPU commands drained per round. Must scale with the
+    // per-round instruction volume: with the superblock runner a single
+    // round legitimately retires up to ~SUPERBLOCK_INSTR_BUDGET per slot
+    // (vs ~6 for the old one-block path), so the rate `executed/6` is much
+    // higher and a flat cap of 64 throttled GPU command processing ~17×
+    // (packets 50279→1861 @50M) — collapsing the present loop / splash.
+    // Cap at the budget so the GPU keeps pace with the CPU at the same
+    // per-instruction rate the one-block path had. The inner loop already
+    // early-breaks on `!gpu.is_ready`, so this only bounds a pathological
+    // backlog, never busy-spins.
+    let gpu_cap = superblock_budget().max(64);
+    if gpu_runs > gpu_cap {
+        gpu_runs = gpu_cap;
     }
     if let Some(gpu) = kernel.gpu.as_inline_mut() {
         gpu.sync_with_mmio();
@@ -2327,10 +2355,22 @@ fn coord_post_round(
     }
 
     if kernel.gpu.has_pending_interrupts() {
-        for _pi in kernel.gpu.take_pending_interrupts() {
+        for pi in kernel.gpu.take_pending_interrupts() {
+            // Canary `ExecutePacketType3_INTERRUPT` dispatches the callback
+            // once per set bit of `cpu_mask` with that bit's index as the
+            // target CPU (`DispatchInterruptCallback(1, n)`). The guest's
+            // swap-acknowledge fence stores `cpu_mask`, and the ISR clears
+            // `1 << current_cpu` from it — so the ISR must run impersonating
+            // the masked CPU or the fence never reaches 0. Sylpheed uses a
+            // single-bit mask (`0x4` → CPU 2); take the lowest set bit.
+            let cpu = if pi.cpu_mask == 0 {
+                xenia_kernel::interrupts::VSYNC_TARGET_CPU
+            } else {
+                pi.cpu_mask.trailing_zeros().min(5) as u8
+            };
             kernel
                 .interrupts
-                .queue_interrupt(xenia_kernel::INTERRUPT_SOURCE_CP);
+                .queue_interrupt(xenia_kernel::INTERRUPT_SOURCE_CP, cpu);
         }
     }
 
@@ -2430,10 +2470,19 @@ fn worker_prologue(
     // and println one record. Read-only; lockstep digest unaffected.
     // Empty set is the common case → single `is_empty()` test inside
     // the helper, no overhead on the hot path.
+    // Perf (Tier-A #3): all four `fire_*_if_match` helpers early-return
+    // on an empty registry, but paying 4× call overhead per slot-visit
+    // (~3.2M visits boot-to-splash) is itself measurable. Gate the whole
+    // group behind a single `any_probe_active()` predicted branch so the
+    // common (no-probe) headless path never even makes the calls. When a
+    // probe IS configured each helper still re-checks its own set, so
+    // behaviour is identical either way.
+    if kernel.any_probe_active() {
         kernel.fire_ctor_probe_if_match(hw_id, mem);
         kernel.fire_branch_probe_if_match(hw_id);
         kernel.fire_audit_pc_probe_if_match(hw_id, mem);
         kernel.fire_lr_trace_if_match(hw_id);
+    }
 
     if mem.has_mem_watch() {
         let ctx = kernel.scheduler.ctx(hw_id);
@@ -2499,8 +2548,15 @@ fn worker_prologue(
         return PrologueOutcome::Continue;
     }
 
-    // 2) Import thunk intercept.
-    if let Some((module, ordinal, name)) = thunk_map.get(&pc) {
+    // 2) Import thunk intercept. Perf (Tier-A #4): import thunks occupy a
+    // small contiguous address band; the overwhelming majority of executing
+    // PCs are ordinary guest code outside it. Range-reject against the band
+    // (two integer compares) before paying the `thunk_map` hash. Faithful
+    // no-op — any in-band PC still goes through the exact map lookup, and an
+    // out-of-band PC can never be a registered thunk.
+    if kernel.pc_in_thunk_band(pc)
+        && let Some((module, ordinal, name)) = thunk_map.get(&pc)
+    {
         let module = *module;
         let ordinal_u32 = *ordinal as u32;
         let thunk_pc = pc;
@@ -2767,6 +2823,160 @@ fn worker_epilogue(
     SlotOutcome::Continue
 }
 
+/// Hard cap on the number of guest instructions a single superblock
+/// runner invocation executes before returning to the round-robin
+/// scheduler. Bounds how coarse the lockstep interleaving can get: a
+/// larger budget amortizes more per-round/per-slot tax (faster) but
+/// runs one HW thread for longer between scheduler returns (coarser
+/// cross-thread interleaving). 1024 keeps a slot-visit ~170× longer
+/// than the old single-block (~6 instr) granularity while still
+/// returning to the round well inside a single 50k quantum. Purely an
+/// instruction count → deterministic, schedule reproduces byte-identically.
+///
+/// Tuned empirically on the Sylpheed boot-to-splash workload (iterate-3AL):
+/// budgets up to 256 keep boot progression byte-for-byte healthy (draws /
+/// swaps / packets track the one-block baseline), then a sharp cliff at
+/// ~384 collapses the present loop (a producer/consumer boot handoff
+/// starves when one slot runs too long without returning to the round).
+/// 128 sits 3× below that cliff with ~1.65× boot-to-splash speedup — a
+/// deliberately conservative pick (correctness over the last few %). The
+/// `XENIA_SUPERBLOCK_BUDGET` env var overrides it for further tuning.
+const SUPERBLOCK_INSTR_BUDGET: u64 = 128;
+
+/// Effective superblock budget. Defaults to [`SUPERBLOCK_INSTR_BUDGET`];
+/// `XENIA_SUPERBLOCK_BUDGET` overrides it (A/B tuning without a rebuild).
+/// A budget of 1 reproduces the old one-block-per-slot-visit behaviour
+/// (the chain always stops after the first block). Read once and cached.
+fn superblock_budget() -> u64 {
+    use std::sync::OnceLock;
+    static BUDGET: OnceLock<u64> = OnceLock::new();
+    *BUDGET.get_or_init(|| {
+        std::env::var("XENIA_SUPERBLOCK_BUDGET")
+            .ok()
+            .and_then(|v| v.parse::<u64>().ok())
+            .filter(|&v| v >= 1)
+            .unwrap_or(SUPERBLOCK_INSTR_BUDGET)
+    })
+}
+
+/// Superblock runner (iterate-3AL). Executes a *chain* of basic blocks
+/// for one slot-visit — following each block's terminating branch into
+/// the next block — instead of a single block, amortizing the per-round
+/// (timebase / coord / `round_schedule`) and per-slot (`worker_prologue`)
+/// dispatch tax over up to [`SUPERBLOCK_INSTR_BUDGET`] guest instructions.
+///
+/// Determinism + cross-thread correctness: the chain ENDS (returns to the
+/// round) at exactly the points where lockstep granularity matters, all
+/// pure functions of guest state (never wall-clock):
+///   - a non-`Continue` step result (Yield / SystemCall / Trap / Unimpl /
+///     Halted) — `step_block` already bails on these; `Yield` in
+///     particular is the db16cyc spin-wait hand-off that prevents a
+///     spinner from starving its producer.
+///   - the just-run block was `sync_sensitive` (reserved load/store or a
+///     memory barrier) — the guest's own ordering points.
+///   - the block touched MMIO (the `mem.mmio_access_count()` watermark
+///     advanced) — GPU/register ordering vs other HW threads stays at the
+///     same fine granularity as the old one-block path.
+///   - the next PC leaves ordinary guest code: an import thunk, the halt
+///     sentinel, or unmapped memory — those need the full `worker_prologue`
+///     dispatch, so we stop and let the next round's prologue handle them.
+///   - the instruction budget is reached.
+///
+/// Instruction-count / clock accounting stays exact: `executed` is summed
+/// from the per-block `cycle_count` delta across every chained block and
+/// handed to `worker_epilogue` once, which advances `stats.instruction_count`
+/// and `decrement_quantum` by precisely the retired count — identical to
+/// dispatching each block separately.
+#[allow(clippy::too_many_arguments)]
+fn run_superblock(
+    wc: &mut WorkerCtx,
+    kernel: &mut xenia_kernel::KernelState,
+    mem: &xenia_memory::GuestMemory,
+    debugger: &mut xenia_debugger::Debugger,
+    thunk_map: &HashMap<u32, (ModuleId, u16, String)>,
+    stats: &mut ExecStats,
+    tid: Option<u32>,
+    thread_ref: xenia_cpu::ThreadRef,
+    first_block_ptr: *const xenia_cpu::block_cache::DecodedBlock,
+    first_pc_before: u32,
+) -> SlotOutcome {
+    use xenia_cpu::interpreter::{step_block, StepResult};
+    const LR_HALT: u32 = xenia_cpu::context::LR_HALT_SENTINEL as u32;
+
+    let budget = superblock_budget();
+
+    // Probe / mem-watch / debugger-hook modes need per-block-entry
+    // observability; in those modes never chain (run exactly one block,
+    // identical to the pre-superblock behaviour). The block-cache fast
+    // path is only entered when hooks/DB are off anyway, but a probe or
+    // mem-watch can be armed alongside it.
+    let chain_allowed = !kernel.any_probe_active() && !mem.has_mem_watch();
+
+    let mut block_ptr = first_block_ptr;
+    let mut pc_before = first_pc_before;
+    let mut total_executed: u64 = 0;
+
+    let (result, last_block_ptr, last_pc_before) = loop {
+        let cycle_before = kernel.scheduler.ctx_mut_ref(thread_ref).cycle_count;
+        let mmio_before = mem.mmio_access_count();
+        let block = unsafe { &*block_ptr };
+        let result = {
+            let ctx = kernel.scheduler.ctx_mut_ref(thread_ref);
+            step_block(ctx, mem, block)
+        };
+        let executed = kernel
+            .scheduler
+            .ctx_mut_ref(thread_ref)
+            .cycle_count
+            .saturating_sub(cycle_before);
+        total_executed = total_executed.saturating_add(executed);
+
+        // STOP conditions (any → end the superblock, hand to epilogue):
+        // non-Continue result (let the epilogue apply it), chaining
+        // disabled, a sync-sensitive block just ran, MMIO was touched,
+        // or the budget is spent.
+        if !chain_allowed
+            || !matches!(result, StepResult::Continue)
+            || block.sync_sensitive
+            || mem.mmio_access_count() != mmio_before
+            || total_executed >= budget
+        {
+            break (result, block_ptr, pc_before);
+        }
+
+        // Decide whether the NEXT PC is an ordinary guest block we can
+        // chain into. Anything else (thunk / halt sentinel / unmapped)
+        // needs the full prologue dispatch next round.
+        let next_pc = kernel.scheduler.ctx(wc.hw_id).pc;
+        if next_pc == LR_HALT
+            || (kernel.pc_in_thunk_band(next_pc) && thunk_map.contains_key(&next_pc))
+            || !mem.is_mapped(next_pc)
+        {
+            break (result, block_ptr, pc_before);
+        }
+
+        // Chain: build/fetch the next block. Re-borrows `wc.block_cache`,
+        // which invalidates the previous `block_ptr` — but we've already
+        // finished using it (only `sync_sensitive`/diagnostics were read,
+        // above), so the raw-pointer aliasing rule is respected.
+        pc_before = next_pc;
+        block_ptr = wc.block_cache.lookup_or_build(next_pc, mem) as *const _;
+    };
+
+    worker_epilogue(
+        wc,
+        kernel,
+        debugger,
+        stats,
+        tid,
+        thread_ref,
+        last_block_ptr,
+        last_pc_before,
+        result,
+        total_executed,
+    )
+}
+
 #[instrument(skip_all, fields(max = ?max_instructions, ips = ?ips_limit))]
 fn run_execution(
     mem: &xenia_memory::GuestMemory,
@@ -2780,8 +2990,6 @@ fn run_execution(
     halt_on_deadlock: bool,
     shutdown: Option<std::sync::Arc<std::sync::atomic::AtomicBool>>,
 ) -> ExecStats {
-    use xenia_cpu::interpreter::step_block;
-
     let mut stats = ExecStats::default();
     let _ = quiet; // retained for future per-kind suppression
 
@@ -2825,6 +3033,10 @@ fn run_execution(
     // re-decoding the same handful of pages 60×/s.
     let mut isr_decode_cache = xenia_cpu::decoder::DecodeCache::new();
 
+    // Tier-A perf #2: reusable buffer for `round_schedule_into` so the round
+    // loop doesn't heap-allocate a `Vec<u8>` every iteration.
+    let mut order_buf = [0u8; xenia_cpu::scheduler::HW_THREAD_COUNT];
+
     'outer: loop {
         // Per-round prologue: budget / shutdown / heartbeat / vsync /
         // timers / audio-interrupt injection. Carved into
@@ -2879,10 +3091,12 @@ fn run_execution(
             thunk_map,
         );
 
-        // Snapshot round schedule. `round_schedule` also advances rng state
-        // when seeded; mutation is intentional.
+        // Snapshot round schedule. `round_schedule_into` also advances rng
+        // state when seeded; mutation is intentional. Perf (Tier-A #2): fill
+        // a reusable stack array instead of allocating a fresh Vec per round.
         kernel.scheduler.begin_round();
-        let order = kernel.scheduler.round_schedule();
+        let order_n = kernel.scheduler.round_schedule_into(&mut order_buf);
+        let order = &order_buf[..order_n];
 
         if order.is_empty() {
             // No Ready threads — advance time to the earliest pending
@@ -2904,7 +3118,7 @@ fn run_execution(
         // GPU when block dispatch engages.
         let instrs_at_round_start = stats.instruction_count;
 
-        for hw_id in order {
+        for &hw_id in order {
             let wc = &mut workers[hw_id as usize];
             match worker_prologue(
                 wc,
@@ -2923,34 +3137,25 @@ fn run_execution(
                     block_ptr,
                     pc_before,
                 } => {
-                    // Block-cache step. The lockstep path keeps the
-                    // kernel state borrowed straight through (single
-                    // host thread, no contention). Step 03 of the
-                    // M3 real-parallelism plan introduces a
-                    // drop-and-reacquire window around `step_block`
-                    // for the parallel branch.
-                    let cycle_before = kernel.scheduler.ctx_mut_ref(thread_ref).cycle_count;
-                    let block = unsafe { &*block_ptr };
-                    let result = {
-                        let ctx = kernel.scheduler.ctx_mut_ref(thread_ref);
-                        step_block(ctx, mem, block)
-                    };
-                    let executed = kernel
-                        .scheduler
-                        .ctx_mut_ref(thread_ref)
-                        .cycle_count
-                        .saturating_sub(cycle_before);
-                    match worker_epilogue(
+                    // SUPERBLOCK runner (iterate-3AL). Instead of one
+                    // basic block per slot-visit, chain straight-line
+                    // blocks through their branches up to a deterministic
+                    // instruction budget, yielding back to the round only
+                    // at cross-thread synchronization points. Amortizes
+                    // the per-round (timebase / coord / round_schedule)
+                    // and per-slot (prologue) tax over hundreds of
+                    // instructions instead of ~6. See `run_superblock`.
+                    match run_superblock(
                         wc,
                         kernel,
+                        mem,
                         debugger,
+                        thunk_map,
                         &mut stats,
                         tid,
                         thread_ref,
                         block_ptr,
                         pc_before,
-                        result,
-                        executed,
                     ) {
                         SlotOutcome::Continue => continue,
                         SlotOutcome::BreakOuter => break 'outer,
@@ -3534,7 +3739,17 @@ fn dispatch_graphics_interrupts(
         None
     };
 
+    /// X_KPCR offset of `prcb_data.current_cpu` (canary `xthread.cc`
+    /// `SetActiveCpu` → `pcr.prcb_data.current_cpu`). The guest graphics
+    /// ISR reads it via `lbz r10, 268(r13)` to decide which per-CPU bit of
+    /// the swap-acknowledge fence to clear.
+    const PCR_CURRENT_CPU_OFF: u32 = 268;
+
     while let Some(source) = kernel.interrupts.peek_next() {
+        let target_cpu = kernel
+            .interrupts
+            .peek_next_cpu()
+            .unwrap_or(xenia_kernel::interrupts::VSYNC_TARGET_CPU);
         // Victim selection: Ready first, then Blocked (canary's
         // `XThread::GetCurrentThread()` analog — any live thread will
         // do for borrowing context). Skip Idle/Exited/ServicingIrq.
@@ -3604,6 +3819,19 @@ fn dispatch_graphics_interrupts(
             saved
         };
 
+        // Impersonate the interrupt's target CPU on the borrowed thread's
+        // PCR, mirroring canary `EmulateCPInterruptDPC` →
+        // `XThread::SetActiveCpu(cpu)`. The guest swap-complete ISR clears
+        // `1 << [pcr.current_cpu]` from the per-present swap-acknowledge
+        // fence; if it runs on the wrong CPU it clears the wrong bit and
+        // the GPU's trailing `WAIT_REG_MEM` on that fence never releases —
+        // stranding the present/title loop. Save/restore so borrowing a
+        // thread doesn't permanently rewrite its processor number.
+        let pcr_addr = (kernel.scheduler.ctx_mut_ref(target_ref).gpr[13] as u32)
+            .wrapping_add(PCR_CURRENT_CPU_OFF);
+        let saved_cpu = mem.read_u8(pcr_addr);
+        mem.write_u8(pcr_addr, target_cpu);
+
         // Stash the previous `scheduler.current` (call_export reaches
         // it; imports the ISR calls must dispatch on the borrowed
         // thread). Restore on the way out.
@@ -3696,6 +3924,7 @@ fn dispatch_graphics_interrupts(
 
         // Restore the borrowed context.
         saved.restore(kernel.scheduler.ctx_mut_ref(target_ref));
+        mem.write_u8(pcr_addr, saved_cpu);
         kernel.scheduler.current = prev_current;
         kernel.interrupts.delivered += 1;
 
@@ -4376,6 +4605,12 @@ fn run_with_ui(
         .map_err(|e| anyhow::anyhow!("winit event loop build failed: {e}"))?;
     let (ui_handles, kernel_bridge) = xenia_ui::build(event_loop.create_proxy());
     kernel.ui = Some(kernel_bridge);
+    // iterate-3O: enable per-draw geometry capture so the UI can replay real
+    // guest draws. Only on the `--ui` path; headless `check` never gets here,
+    // so the deterministic core/golden stays untouched.
+    if let Some(gpu) = kernel.gpu.as_inline_mut() {
+        gpu.enable_frame_capture();
+    }
 
     let shutdown = std::sync::Arc::clone(&ui_handles.shutdown);
     let title_owned = std::path::Path::new(title)

crates/xenia-app/tests/golden/sylpheed_n2m.json

@@ -1,5 +1,5 @@
 {
-  "instructions": 2000005,
+  "instructions": 2000073,
   "imports": 5635,
   "unimpl": 0,
   "draws": 0,

crates/xenia-app/tests/golden/sylpheed_n50m.json

@@ -1,10 +1,10 @@
 {
-  "instructions": 50000003,
-  "imports": 451508,
+  "instructions": 50000110,
+  "imports": 243387,
   "unimpl": 0,
-  "draws": 0,
-  "swaps": 2,
-  "unique_render_targets": 0,
-  "shader_blobs_live": 0,
-  "texture_cache_entries": 0
+  "draws": 1279,
+  "swaps": 260,
+  "unique_render_targets": 2,
+  "shader_blobs_live": 6,
+  "texture_cache_entries": 1
 }

crates/xenia-app/tests/sylpheed_oracles.rs

@@ -57,6 +57,16 @@ fn run_oracle(label: &str, max_instr: u64, golden_rel: &str) {
             &iso,
             "-n",
             &max_instr_str,
+            // Pin the inline (single-threaded) GPU backend. The default
+            // threaded backend drains the ring on a separate host thread,
+            // so the exact instruction at which a CP interrupt is queued —
+            // and therefore when the guest's swap-complete ISR callback runs
+            // (iterate-2S armed it via SCRATCH_REG writeback) — varies run to
+            // run. Inline draining is instruction-count-deterministic, which
+            // is what a regression golden needs. (The threaded path is the
+            // documented "GPU thread race" the stable-digest already warns
+            // about.)
+            "--gpu-inline",
             "--stable-digest",
             "--expect",
             &golden_str,

crates/xenia-cpu/src/block_cache.rs

@@ -79,6 +79,14 @@ pub struct DecodedBlock {
     /// a successful build (`MAX_BLOCK_INSTRS >= 1` and the build walk
     /// pushes the first decoded word unconditionally).
     pub instrs: Vec<DecodedInstr>,
+    /// True if this block contains a cross-thread synchronization point
+    /// (`PpcOpcode::is_sync_sensitive`: reserved load/store or a memory
+    /// barrier). Computed once at build time. The superblock runner ends
+    /// the run after executing a sync-sensitive block so the lockstep
+    /// interleaving stays fine-grained at exactly those points (preserving
+    /// the cross-thread ordering the 2E/2F/2J boot work depends on),
+    /// while chaining freely through ordinary straight-line blocks.
+    pub sync_sensitive: bool,
 }
 
 /// Per-slot status from a `lookup_or_build` probe. Internal only.
@@ -187,11 +195,13 @@ fn build_block(start_pc: u32, mem: &dyn MemoryAccess, page_version: u64) -> Deco
     let mut instrs: Vec<DecodedInstr> = Vec::with_capacity(8);
     let page_base = start_pc & GUEST_PAGE_MASK;
     let mut cur = start_pc;
+    let mut sync_sensitive = false;
 
     loop {
         let raw = mem.read_u32(cur);
         let decoded = decode(raw, cur);
         let terminates = decoded.opcode.terminates_block();
+        sync_sensitive |= decoded.opcode.is_sync_sensitive();
         instrs.push(decoded);
 
         if terminates {
@@ -215,6 +225,7 @@ fn build_block(start_pc: u32, mem: &dyn MemoryAccess, page_version: u64) -> Deco
         end_pc,
         page_version,
         instrs,
+        sync_sensitive,
     }
 }
 
@@ -335,6 +346,40 @@ mod tests {
         assert_eq!(b.end_pc, 0x110);
     }
 
+    #[test]
+    fn sync_sensitive_flag_set_for_barrier_block() {
+        // A block containing `sync` (0x7C0004AC) must flag sync_sensitive
+        // so the superblock runner ends the chain there (cross-thread
+        // ordering point). `sync` does NOT terminate a block, so it sits
+        // mid-block followed by straight-line code up to a terminator.
+        let mem = BlockTestMem::new();
+        mem.put(0x100, enc_addi(3, 3, 1));
+        mem.put(0x104, 0x7C00_04AC); // sync
+        mem.put(0x108, enc_addi(3, 3, 1));
+        mem.put(0x10C, enc_b_self()); // terminator
+        let mut bc = BlockCache::new();
+        let b = bc.lookup_or_build(0x100, &mem);
+        assert!(
+            b.sync_sensitive,
+            "block containing `sync` must flag sync_sensitive; decoded last={:?}",
+            b.instrs.iter().map(|i| i.opcode).collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn sync_sensitive_flag_clear_for_plain_block() {
+        // A straight-line ALU block with no reserved-op / barrier must
+        // NOT flag sync_sensitive (so the superblock runner is free to
+        // chain through it).
+        let mem = BlockTestMem::new();
+        mem.put(0x100, enc_addi(3, 3, 1));
+        mem.put(0x104, enc_addi(3, 3, 1));
+        mem.put(0x108, enc_b_self());
+        let mut bc = BlockCache::new();
+        let b = bc.lookup_or_build(0x100, &mem);
+        assert!(!b.sync_sensitive, "plain ALU block must not flag sync_sensitive");
+    }
+
     #[test]
     fn block_stops_at_page_boundary() {
         // Build from 0x1FFC. The next PC (0x2000) is in a different

crates/xenia-cpu/src/opcode.rs

@@ -204,6 +204,34 @@ impl PpcOpcode {
         )
     }
 
+    /// Returns true if this opcode is a cross-thread synchronization
+    /// point at which the superblock runner MUST yield back to the
+    /// round-robin scheduler so the lockstep interleaving stays
+    /// fine-grained enough to preserve correct cross-thread ordering:
+    ///
+    ///   - reserved load/store (`lwarx`/`ldarx`/`stwcx.`/`stdcx.`): the
+    ///     atomic primitive other threads race on. Running past one
+    ///     without returning to the scheduler would let a single slot
+    ///     win/lose a reservation across many blocks before any peer
+    ///     observes it.
+    ///   - memory barriers (`sync`/`eieio`/`isync`): the guest explicitly
+    ///     demands a global ordering point here; honour it by ending the
+    ///     superblock so the scheduler re-interleaves.
+    ///
+    /// Purely a function of the opcode (no guest data), so the yield
+    /// decision is deterministic and the schedule reproduces byte-identically.
+    /// Note: `sc` (syscall) and traps already `terminates_block`, and
+    /// import-thunk / halt-sentinel PCs are handled by the per-block
+    /// prologue re-check in the superblock loop — they are not listed here.
+    #[inline]
+    pub fn is_sync_sensitive(&self) -> bool {
+        matches!(
+            self,
+            Self::lwarx | Self::ldarx | Self::stwcx | Self::stdcx
+                | Self::sync | Self::eieio | Self::isync
+        )
+    }
+
     pub fn name(&self) -> &'static str {
         match self {
             Self::Invalid => "invalid",

crates/xenia-cpu/src/scheduler.rs

@@ -795,31 +795,46 @@ impl Scheduler {
     /// the fast path — zero bits mean no slot has work and the caller
     /// falls through to `advance_to_next_wake`.
     pub fn round_schedule(&mut self) -> Vec<u8> {
+        let mut buf = [0u8; HW_THREAD_COUNT];
+        let n = self.round_schedule_into(&mut buf);
+        buf[..n].to_vec()
+    }
+
+    /// Allocation-free variant of [`Self::round_schedule`] (Tier-A perf #2).
+    /// Fills `buf` with the runnable slot ids and returns the count `n`; the
+    /// valid range is `buf[..n]`. The hot scheduler loop (lockstep +
+    /// parallel) calls this with a reusable stack array so it does not
+    /// `__rust_alloc`/`__rust_dealloc` a fresh `Vec` every round (~7 instr
+    /// apart at boot-to-splash → millions of churned allocations). Identical
+    /// ordering / RNG-advance semantics to `round_schedule`, so the schedule
+    /// — and thus the lockstep digest — is byte-for-byte unchanged.
+    pub fn round_schedule_into(&mut self, buf: &mut [u8; HW_THREAD_COUNT]) -> usize {
         if self.non_empty_runnable == 0 {
-            return Vec::new();
+            return 0;
         }
         let start = self.rotation_cursor as usize;
-        let mut out: Vec<u8> = Vec::with_capacity(HW_THREAD_COUNT);
+        let mut n = 0usize;
         for off in 0..HW_THREAD_COUNT {
             let i = (start + off) % HW_THREAD_COUNT;
             if self.non_empty_runnable & (1 << i) != 0 {
-                out.push(i as u8);
+                buf[n] = i as u8;
+                n += 1;
             }
         }
         // Seeded mode layers a deterministic shuffle on top of the
         // already-filtered list. Same spawn/wake sequence + same seed ⇒
         // same schedule (invariant preserved from pre-Axis-1).
         if let OrderMode::Seeded { .. } = self.order {
-            for i in (1..out.len()).rev() {
+            for i in (1..n).rev() {
                 self.rng_state ^= self.rng_state << 13;
                 self.rng_state ^= self.rng_state >> 7;
                 self.rng_state ^= self.rng_state << 17;
                 let j = (self.rng_state as usize) % (i + 1);
-                out.swap(i, j);
+                buf.swap(i, j);
             }
         }
         self.rotation_cursor = ((start + 1) % HW_THREAD_COUNT) as u8;
-        out
+        n
     }
 
     pub fn begin_round(&mut self) {
@@ -1293,7 +1308,15 @@ impl Scheduler {
         };
         t.quantum_remaining = QUANTUM_DEFAULT;
         self.recompute_slot_runnable(r.hw_id);
-        tracing::info!(
+        // DEBUG, not INFO: this fires once per timed-wait deadline-wake, which
+        // during the boot idle-spin happens hundreds of thousands of times. At
+        // INFO it floods the console/log file and throttles the interactive
+        // `exec --ui` path so hard (≈286K lines flushed to disk) that the guest
+        // crawls and never reaches the ~30–150M-instruction splash window —
+        // which masqueraded as a "--ui early termination" (iterate-3R). The
+        // headless `check` path runs `--quiet` (WARN) so it was never throttled.
+        // No execution-semantics change; deterministic golden is unaffected.
+        tracing::debug!(
             "scheduler: advanced to deadline {} waking hw={} idx={}",
             deadline,
             r.hw_id,

crates/xenia-gpu/src/draw_capture.rs

@@ -0,0 +1,372 @@
+//! Per-draw geometry capture for the host UI's faithful-render path.
+//!
+//! The deterministic headless core (`check --gpu-inline`) never touches this
+//! module — it is populated only when a UI bridge is installed and consumed
+//! only by `crates/xenia-ui`. The goal is to hand the UI the *real* guest
+//! geometry behind each `PM4_DRAW_INDX*` packet so it can rasterize the
+//! actual splash vertices instead of synthetic placeholder shapes.
+//!
+//! What the WGSL pipeline needs to reconstruct one draw (see
+//! `shaders/xenos_interp.wgsl` `vs_main` / `interpret_vertex_fetch`):
+//!   * the active VS/PS blob keys (already published as assets),
+//!   * the primitive type + the host vertex count to issue,
+//!   * the raw guest vertex-buffer bytes for the fetched window, and
+//!   * the *dword base* of that window so the shader can rebase the absolute
+//!     fetch-constant address into the uploaded buffer.
+//!
+//! The hard part is sourcing the vertex window: the VS reads a vertex-fetch
+//! constant (`xe_gpu_vertex_fetch_t`) whose dword-0 carries the absolute
+//! guest dword address. We parse the active VS, find its first vertex fetch,
+//! read that fetch constant out of the register file, then copy a bounded
+//! window of guest memory starting at the fetch base.
+
+use xenia_memory::access::MemoryAccess;
+
+use crate::draw_state::{IndexSize, IndexSource, PrimitiveType};
+use crate::register_file::RegisterFile;
+
+/// Texture-fetch / vertex-fetch constant region base, in register indices.
+/// Each fetch constant is 6 dwords (`xe_gpu_*_fetch_t`).
+const CONST_BASE_FETCH: u32 = 0x4800;
+
+/// Upper bound (in dwords) on the vertex window we copy per draw. The splash
+/// UI draws are tiny (3–4 verts × ≤4 dwords); 64 KiB of dwords is generous
+/// slack while bounding the per-frame copy cost and the 16 MiB host buffer.
+const MAX_WINDOW_DWORDS: u32 = 16 * 1024;
+
+/// One captured draw, with enough real state for the UI to replay it through
+/// the existing wgpu Xenos pipeline.
+#[derive(Clone, Debug)]
+pub struct DrawCapture {
+    /// Monotonic global draw index (matches `GpuStats::draws_seen` at capture).
+    pub draw_index: u32,
+    /// Xenos primitive-type code (see `SwapInfo::last_draw_prim` encoding).
+    pub prim_code: u32,
+    /// Host vertex count to issue (post primitive-processor rewrite).
+    pub host_vertex_count: u32,
+    /// Active VS blob key at draw time (0 = none).
+    pub vs_key: u32,
+    /// Active PS blob key at draw time (0 = none).
+    pub ps_key: u32,
+    /// Raw guest dwords of the fetched vertex window (host-endian as stored in
+    /// guest memory — the WGSL applies the per-format endian swap). `addr 0`
+    /// of this buffer corresponds to guest dword `window_base_dwords`.
+    pub vertex_dwords: Vec<u32>,
+    /// Guest dword address that maps to index 0 of `vertex_dwords`. The shader
+    /// subtracts this from the fetch-constant base to index `vertex_dwords`.
+    pub window_base_dwords: u32,
+    /// `true` when we successfully resolved a real vertex window. When `false`
+    /// the UI falls back to its procedural geometry for this draw (honest:
+    /// nothing faked, just "couldn't source real vertices").
+    pub has_real_vertices: bool,
+    /// iterate-3S: per-draw NDC transform derived from the guest viewport /
+    /// clip / VTE registers (mirrors canary `GetHostViewportInfo`). The host VS
+    /// converts the guest-VS position to wgpu clip space via
+    /// `clip.xy = pos.xy * ndc_scale + ndc_offset * pos.w`. The Y component
+    /// already carries the render-target → wgpu Y-flip (negated).
+    pub ndc_scale: [f32; 2],
+    pub ndc_offset: [f32; 2],
+    /// iterate-3T: the decoded texture(s) this draw's active pixel shader
+    /// samples, keyed off its real `tfetch` fetch-constant slots (the 3M
+    /// decoder makes these decode). The UI uploads + binds the FIRST entry
+    /// per-draw so the textured logo samples the real artwork instead of the
+    /// magenta stub. Empty for flat (no-tfetch) draws. Populated by
+    /// `gpu_system` after decode (left empty by `build`).
+    ///
+    /// Each entry is `(key, content_version, bytes)`. iterate-3AD: the
+    /// `content_version` (from `span_max_version` over the texel span) lets the
+    /// UI host texture cache RE-UPLOAD when the guest fills more of an evolving
+    /// atlas. The publisher and the 2nd splash logo share one K8888 surface
+    /// (base `0x4dbee000`); the 2nd logo's texels are CPU-written *after* the
+    /// publisher's first upload. Without the real version the host cache (which
+    /// previously pinned `version_when_uploaded = 1`) kept the first partial
+    /// upload, so the 2nd logo sampled its still-zero atlas region as black.
+    pub textures: Vec<(crate::texture_cache::TextureKey, u64, Vec<u8>)>,
+    /// iterate-3Y: per-draw color/blend render state captured from the
+    /// register file so the host pipeline composites the way the guest
+    /// intends (instead of one fixed alpha-blend state). Mirrors the fields
+    /// canary feeds into `GetCurrentStateDescription` (D3D12
+    /// `pipeline_cache.cc`):
+    ///   * `blend_control` = `RB_BLENDCONTROL0` (RT0 src/dst factors + op,
+    ///     color and alpha). The Xbox 360 has no separate "blend enable" bit;
+    ///     `One,Zero,Add` *is* the opaque case.
+    ///   * `color_mask` = RT0 nibble of `RB_COLOR_MASK` (per-channel write
+    ///     enable). When 0, canary forces `One,Zero` (no blend).
+    ///   * `color_control` = `RB_COLORCONTROL` (alpha-test enable/func).
+    ///   * `depth_control` = `RB_DEPTHCONTROL` (z-test enable/func/write).
+    pub blend_control: u32,
+    pub color_mask: u8,
+    pub color_control: u32,
+    pub depth_control: u32,
+}
+
+/// iterate-3S: compute the guest→host NDC XY transform for a draw, mirroring
+/// canary's `draw_util.cc::GetHostViewportInfo` (the XY half). The Xbox 360 VS
+/// emits a clip-space position which the HW then scales/offsets by the viewport
+/// (`PA_CL_VPORT_*`, gated by `PA_CL_VTE_CNTL`) into render-target pixels, OR,
+/// when clipping is disabled (`PA_CL_CLIP_CNTL.clip_disable`), the VS emits
+/// render-target-pixel coordinates directly (the screen-space UI / clear case —
+/// this is what Sylpheed's splash quads do). Either way we must rescale into the
+/// host's [-1,1] clip space and flip Y (render-target Y-down → wgpu Y-up).
+///
+/// Returns `(ndc_scale[2], ndc_offset[2])` such that
+/// `host_clip.xy = guest_pos.xy * ndc_scale + ndc_offset * guest_pos.w`.
+/// The Y entries are pre-negated to flip into wgpu's Y-up clip space.
+pub fn compute_ndc_xy(rf: &RegisterFile) -> ([f32; 2], [f32; 2]) {
+    const PA_CL_CLIP_CNTL: u32 = 0x2204;
+    const PA_SU_SC_MODE_CNTL: u32 = 0x2205;
+    const PA_CL_VTE_CNTL: u32 = 0x2206;
+    const PA_SU_VTX_CNTL: u32 = 0x2302;
+    const PA_CL_VPORT_XSCALE: u32 = 0x210F;
+    const PA_CL_VPORT_XOFFSET: u32 = 0x2110;
+    const PA_CL_VPORT_YSCALE: u32 = 0x2111;
+    const PA_CL_VPORT_YOFFSET: u32 = 0x2112;
+    const PA_SC_WINDOW_OFFSET: u32 = 0x2080;
+    const PA_SC_WINDOW_SCISSOR_BR: u32 = 0x2082;
+    const RB_SURFACE_INFO: u32 = 0x2000;
+
+    let clip_cntl = rf.read(PA_CL_CLIP_CNTL);
+    let vte = rf.read(PA_CL_VTE_CNTL);
+    let su_sc_mode = rf.read(PA_SU_SC_MODE_CNTL);
+    let su_vtx = rf.read(PA_SU_VTX_CNTL);
+    let fbits = |r: u32| f32::from_bits(rf.read(r));
+
+    // VTE enable bits (xenos.h PA_CL_VTE_CNTL): bit0 vport_x_scale_ena,
+    // bit1 vport_x_offset_ena, bit2 vport_y_scale_ena, bit3 vport_y_offset_ena.
+    let scale_x = if vte & (1 << 0) != 0 { fbits(PA_CL_VPORT_XSCALE) } else { 1.0 };
+    let off_x = if vte & (1 << 1) != 0 { fbits(PA_CL_VPORT_XOFFSET) } else { 0.0 };
+    let scale_y = if vte & (1 << 2) != 0 { fbits(PA_CL_VPORT_YSCALE) } else { 1.0 };
+    let off_y = if vte & (1 << 3) != 0 { fbits(PA_CL_VPORT_YOFFSET) } else { 0.0 };
+
+    // Render-target extent in guest pixels: clamp to the texture max (2048),
+    // sourced from the window scissor BR (matches canary `x_max`/`y_max`).
+    let br = rf.read(PA_SC_WINDOW_SCISSOR_BR);
+    let x_max = ((br & 0x7FFF).max(1)).min(2048) as f32;
+    let y_max = (((br >> 16) & 0x7FFF).max(1)).min(2048) as f32;
+    let _ = RB_SURFACE_INFO;
+
+    // Half-pixel + window offsets added in render-target pixels.
+    let mut add_x = 0.0f32;
+    let mut add_y = 0.0f32;
+    if su_sc_mode & (1 << 16) != 0 {
+        let wo = rf.read(PA_SC_WINDOW_OFFSET);
+        // 15-bit signed each (x: [14:0], y: [30:16]).
+        let sx = (((wo & 0x7FFF) << 1) as i32) >> 1;
+        let sy = ((((wo >> 16) & 0x7FFF) << 1) as i32) >> 1;
+        add_x += sx as f32;
+        add_y += sy as f32;
+    }
+    if su_vtx & 1 == 0 {
+        // pix_center == kD3DZero → +0.5 half-pixel offset.
+        add_x += 0.5;
+        add_y += 0.5;
+    }
+
+    let (s, o);
+    if clip_cntl & (1 << 16) != 0 {
+        // clip_disable: VS outputs render-target-*pixel* coords (Y-DOWN: pixel
+        // y=0 is the top row of the render target). Rescale the whole RT extent
+        // to [-1,1] and FLIP Y so pixel-top → wgpu clip-top (canary's
+        // huge-host-viewport path; the framebuffer→clip flip is real here).
+        let px2ndc_x = 2.0 / x_max;
+        let px2ndc_y = 2.0 / y_max;
+        let sx = scale_x * px2ndc_x;
+        let ox = (off_x - x_max * 0.5 + add_x) * px2ndc_x;
+        let sy = scale_y * px2ndc_y;
+        let oy = (off_y - y_max * 0.5 + add_y) * px2ndc_y;
+        // Flip Y: pixel-Y-down → wgpu clip-Y-up.
+        s = [sx, -sy];
+        o = [ox, -oy];
+    } else {
+        // iterate-3AA (DEFECT 1 ROOT): clipping enabled → the VS already emits
+        // *clip-space* coordinates (Y-UP: +Y is the top of the screen), exactly
+        // the convention the Xbox 360's D3D9 and wgpu BOTH use for clip space
+        // (NDC +Y → framebuffer top in each API; the framebuffer Y-direction is
+        // an internal viewport detail handled identically by both). A clip-space
+        // position is therefore portable to wgpu with NO Y-flip. The previous
+        // code unconditionally negated Y (the same flip the screen-space pixel
+        // path needs), which mirrored the publisher logo vertically: its quad is
+        // centered (±0.085 around 0) so the *position* stayed centered, but the
+        // negation swapped top↔bottom vertices while the texture V was unchanged
+        // → the sampled sub-rect (UV v 0.001→0.090) read bottom-up → "SQUARE
+        // ENIX" rendered upside down in place. Measured (readback): the red dots
+        // sit at 43% from the texture top but rendered at 58% from the top
+        // (= a clean vertical mirror); removing the flip restores them to 43%.
+        // Identity XY (no flip) maps guest clip-Y-up straight to wgpu clip-Y-up.
+        s = [1.0, 1.0];
+        o = [0.0, 0.0];
+        return (s, o);
+    }
+    (s, o)
+}
+
+/// Encode a [`PrimitiveType`] as the raw Xenos code used across the bridge.
+pub fn prim_code(p: PrimitiveType) -> u32 {
+    match p {
+        PrimitiveType::None => 0,
+        PrimitiveType::PointList => 1,
+        PrimitiveType::LineList => 2,
+        PrimitiveType::LineStrip => 3,
+        PrimitiveType::TriangleList => 4,
+        PrimitiveType::TriangleFan => 5,
+        PrimitiveType::TriangleStrip => 6,
+        PrimitiveType::RectangleList => 8,
+        PrimitiveType::QuadList => 13,
+        PrimitiveType::Unknown(x) => x as u32,
+    }
+}
+
+/// Resolve the first vertex-fetch window referenced by the parsed VS.
+///
+/// Walks the VS instruction stream for the first `vfetch` (mini) instruction,
+/// reads its fetch constant from `rf`, and copies a bounded window of guest
+/// memory starting at the fetch base. Returns `(dwords, window_base_dwords)`
+/// or `None` if the VS has no vertex fetch or the constant is malformed.
+fn resolve_vertex_window(
+    parsed_vs: &crate::ucode::ParsedShader,
+    rf: &RegisterFile,
+    mem: &dyn MemoryAccess,
+) -> Option<(Vec<u32>, u32)> {
+    // iterate-3W (GPUBUG-109): the instruction block packs ALU and fetch
+    // instructions identically (96 bits / 3 dwords each); ONLY the owning
+    // `Exec` control-flow clause's `sequence` bitmap (2 bits per instruction,
+    // bit[2*i]=fetch/ALU) tells them apart. The previous blind triple-walk
+    // decoded ALU triples as fetches → garbage fetch-constant indices and a
+    // bogus `type==3` guard, never reaching the real vertex fetch. Walk the CF
+    // exec clauses exactly as the translator does (`translator.rs::emit_exec`)
+    // and take the FIRST sequence-flagged *vertex* fetch.
+    let instrs = &parsed_vs.instructions;
+    let mut const_off: Option<u32> = None;
+    'clauses: for clause in &parsed_vs.cf {
+        let crate::ucode::control_flow::ControlFlowInstruction::Exec {
+            address,
+            count,
+            sequence,
+            ..
+        } = *clause
+        else {
+            continue;
+        };
+        for i in 0..(count as usize) {
+            // bit[2*i] of the sequence bitmap: 1 = fetch, 0 = ALU.
+            if (sequence >> (i * 2)) & 1 == 0 {
+                continue;
+            }
+            let base = (address as usize + i) * 3;
+            if base + 2 >= instrs.len() {
+                break;
+            }
+            if let crate::ucode::fetch::FetchInstruction::Vertex(vf) =
+                crate::ucode::fetch::decode_fetch([instrs[base], instrs[base + 1], instrs[base + 2]])
+            {
+                const_off = Some(vf.const_reg_offset());
+                break 'clauses;
+            }
+        }
+    }
+    // iterate-3X (GPUBUG-110): vertex fetch constants are addressed by
+    // `const_index * 3 + const_index_sel` (canary `ucode.h:700` —
+    // `VertexFetchInstruction::fetch_constant_index`), NOT by `const_index`
+    // alone. The register region packs 3 two-dword vertex-fetch constants per
+    // 6-dword group, so the constant lives at
+    // `0x4800 + const_index*6 + const_index_sel*2`. The previous decode dropped
+    // `const_index_sel` and read sub-slot 0 (`fc*6`), which for the publisher
+    // logo (`const_index=31, sel=2`) held `0x00000001` (an unused slot) instead
+    // of the real vertex-buffer base at sub-slot 2 (`0x48BE`). That made
+    // `has_real_vertices=false` → the logo fell to the procedural fullscreen
+    // magenta fallback. (Refutes iterate-3W's "geometry is auto-generated from
+    // vertex_id" — measured: the real fetch constant is a 4-vertex QuadList
+    // buffer at `0x0adf60f0`.)
+    let const_reg = CONST_BASE_FETCH + const_off?;
+    let dword0 = rf.read(const_reg);
+    let dword1 = rf.read(const_reg + 1);
+    // address:30 at bits[31:2] of dword0 (in bytes once masked). The fetch
+    // constant carries a guest *physical* dword address — canary reads the
+    // vertex buffer via `Memory::TranslatePhysical(fetch.address * 4)`
+    // (`draw_util.cc:961`). On the Xbox 360 the physical range is mirrored at
+    // several virtual windows; ours only maps the cached-physical window at
+    // `0x4000_0000` (`gpu_system::physical_to_backing`). Reading the bare low
+    // address (`0x0adf_xxxx`) hits an unmapped VA and returns zeros, so rebase
+    // a low physical base onto the mapped `0x4000_0000` alias when the raw VA
+    // is not itself mapped. `window_base_dwords` keeps the *original* base so
+    // the shader's rebase against the (unmodified) fetch-constant address still
+    // indexes the uploaded window correctly.
+    let base_bytes = dword0 & 0xFFFF_FFFC;
+    if base_bytes == 0 {
+        return None;
+    }
+    let read_base = if mem.translate(base_bytes).is_some() {
+        base_bytes
+    } else if base_bytes < 0x2000_0000 && mem.translate(base_bytes | 0x4000_0000).is_some() {
+        base_bytes | 0x4000_0000
+    } else {
+        base_bytes
+    };
+    // size:24 at bits[25:2] of dword1, in dwords. Clamp to our window cap.
+    let size_dwords = ((dword1 >> 2) & 0x00FF_FFFF).clamp(1, MAX_WINDOW_DWORDS);
+    let window_base_dwords = base_bytes >> 2;
+    let mut dwords = Vec::with_capacity(size_dwords as usize);
+    for i in 0..size_dwords {
+        let addr = read_base.wrapping_add(i * 4);
+        if addr < read_base {
+            break; // wrap guard
+        }
+        // `read_u32` composes big-endian bytes into the u32 value; the WGSL's
+        // `gpu_swap` expects the *raw little-endian dword* as it sits in guest
+        // memory, so undo the BE composition with `swap_bytes`.
+        dwords.push(mem.read_u32(addr).swap_bytes());
+    }
+    if dwords.is_empty() {
+        return None;
+    }
+    Some((dwords, window_base_dwords))
+}
+
+/// Build a [`DrawCapture`] for one draw. Best-effort: when the vertex window
+/// can't be resolved, `has_real_vertices` is `false` and the UI falls back to
+/// procedural geometry (never fabricated pixels).
+#[allow(clippy::too_many_arguments)]
+pub fn build(
+    draw_index: u32,
+    primitive: PrimitiveType,
+    host_vertex_count: u32,
+    _index_source: IndexSource,
+    _index_size: IndexSize,
+    vs_key: u32,
+    ps_key: u32,
+    parsed_vs: Option<&crate::ucode::ParsedShader>,
+    rf: &RegisterFile,
+    mem: &dyn MemoryAccess,
+) -> DrawCapture {
+    let (vertex_dwords, window_base_dwords, has_real) = match parsed_vs
+        .and_then(|vs| resolve_vertex_window(vs, rf, mem))
+    {
+        Some((d, base)) => (d, base, true),
+        None => (Vec::new(), 0, false),
+    };
+    let (ndc_scale, ndc_offset) = compute_ndc_xy(rf);
+    // iterate-3Y: capture RT0 color/blend/depth render state. Registers per
+    // canary `registers.h`: RB_BLENDCONTROL0=0x2201, RB_COLOR_MASK=0x2104
+    // (RT0 = bits[3:0]), RB_COLORCONTROL=0x2202, RB_DEPTHCONTROL=0x2200.
+    const RB_BLENDCONTROL_0: u32 = 0x2201;
+    const RB_COLOR_MASK: u32 = 0x2104;
+    const RB_COLORCONTROL: u32 = 0x2202;
+    const RB_DEPTHCONTROL: u32 = 0x2200;
+    DrawCapture {
+        draw_index,
+        prim_code: prim_code(primitive),
+        host_vertex_count,
+        vs_key,
+        ps_key,
+        vertex_dwords,
+        window_base_dwords,
+        has_real_vertices: has_real,
+        ndc_scale,
+        ndc_offset,
+        textures: Vec::new(),
+        blend_control: rf.read(RB_BLENDCONTROL_0),
+        color_mask: (rf.read(RB_COLOR_MASK) & 0xF) as u8,
+        color_control: rf.read(RB_COLORCONTROL),
+        depth_control: rf.read(RB_DEPTHCONTROL),
+    }
+}

crates/xenia-gpu/src/gpu_system.rs

@@ -78,6 +78,30 @@ pub fn physical_to_backing(addr: u32) -> u32 {
     }
 }
 
+/// Max guest page-version over the `[base, base+len)` span, walking 4 KiB
+/// pages via the `MemoryAccess` trait's `page_version`.
+///
+/// The concrete heap exposes an inherent `max_page_version(base, len)`, but
+/// the draw handler only holds `&dyn MemoryAccess` (which carries the coarser
+/// `page_version(addr)` accessor). This is byte-equivalent to
+/// `heap::max_page_version` and stays a pure function of the per-page write
+/// counters (no wall-clock), so texture-decode timing remains deterministic.
+fn span_max_version(mem: &dyn MemoryAccess, base: u32, len: u32) -> u64 {
+    const PAGE: u32 = 0x1000;
+    let last = base.saturating_add(len.saturating_sub(1));
+    let mut page = base & !(PAGE - 1);
+    let last_page = last & !(PAGE - 1);
+    let mut max = 0u64;
+    loop {
+        max = max.max(mem.page_version(page));
+        if page >= last_page {
+            break;
+        }
+        page = page.wrapping_add(PAGE);
+    }
+    max
+}
+
 /// Cached Xenos microcode blob, produced by `PM4_IM_LOAD*` packets.
 #[derive(Debug, Clone)]
 pub struct ShaderBlob {
@@ -400,12 +424,24 @@ pub struct GpuSystem {
     /// on every texture-fetch resolution; the UI thread sees the decoded
     /// bytes via `UiBridge::publish_texture`.
     pub texture_cache: crate::texture_cache::TextureCache,
+    /// P5b: textures decoded at the most recent `PM4_DRAW_INDX*`, keyed off
+    /// the *active* pixel shader's real `tfetch` fetch-constant slots (not a
+    /// hardcoded slot). `vd_swap` publishes the first of these to the UI so
+    /// the replay binds the texture the draw actually samples. Cleared and
+    /// repopulated each draw; empty when the active PS issues no `tfetch`.
+    pub last_draw_textures: Vec<(crate::texture_cache::TextureKey, u64, Vec<u8>)>,
     /// 10 MiB shadow of the Xenos EDRAM. Written by clear-resolves and
     /// (future) host-render-target readback; read by the resolve byte-copy
     /// path that writes tiled pixels into guest memory. Allocated once at
     /// `GpuSystem::new` and lives for the whole GPU lifetime — no
     /// per-frame churn.
     pub edram: crate::edram::ShadowEdram,
+    /// UI-only: when `Some`, every `PM4_DRAW_INDX*` appends a
+    /// [`crate::draw_capture::DrawCapture`] here so the host UI can replay the
+    /// real guest geometry. `None` in headless/deterministic mode — the
+    /// `--gpu-inline` golden never enables this, so capture is entirely inert
+    /// for `check`. Drained (taken) by `vd_swap` at each present.
+    pub frame_captures: Option<Vec<crate::draw_capture::DrawCapture>>,
 }
 
 impl GpuSystem {
@@ -431,7 +467,17 @@ impl GpuSystem {
             rt_cache: crate::render_target_cache::RenderTargetCache::new(),
             last_resolve: None,
             texture_cache: crate::texture_cache::TextureCache::new(),
+            last_draw_textures: Vec::new(),
             edram: crate::edram::ShadowEdram::new(),
+            frame_captures: None,
+        }
+    }
+
+    /// Enable per-draw geometry capture for the host UI. Inert (and never
+    /// called) in headless/deterministic mode. Idempotent.
+    pub fn enable_frame_capture(&mut self) {
+        if self.frame_captures.is_none() {
+            self.frame_captures = Some(Vec::new());
         }
     }
 
@@ -603,14 +649,21 @@ impl GpuSystem {
     /// Release.
     pub fn sync_with_mmio(&mut self) {
         let wptr_dwords = self.mmio.cp_rb_wptr.load(Ordering::Acquire);
-        if wptr_dwords != self.ring.write_offset_dwords && self.ring.size_dwords != 0 {
-            self.ring.write_offset_dwords = wptr_dwords % self.ring.size_dwords;
+        // CP_RB_WPTR governs ONLY the primary ring. While an indirect buffer
+        // is executing, the active `self.ring` is a fixed linear sub-stream
+        // and the primary ring is saved at the bottom of the IB stack —
+        // applying the (primary) write pointer to the IB would corrupt its
+        // extent (e.g. `wptr % ib_size`) and strand the GPU mid-buffer.
+        let primary = self.ib_stack.first_mut().unwrap_or(&mut self.ring);
+        if wptr_dwords != primary.write_offset_dwords && primary.size_dwords != 0 {
+            primary.write_offset_dwords = wptr_dwords % primary.size_dwords;
         }
-        // Mirror our read pointer (Release pairs with any guest-side
+        let primary_rptr = primary.read_offset_dwords;
+        // Mirror the *primary* read pointer (Release pairs with any guest-side
         // Acquire-load of CP_RB_RPTR for ring writeback bookkeeping).
         self.mmio
             .cp_rb_rptr
-            .store(self.ring.read_offset_dwords, Ordering::Release);
+            .store(primary_rptr, Ordering::Release);
     }
 
     /// True iff `execute_one` is expected to make progress without blocking.
@@ -618,7 +671,11 @@ impl GpuSystem {
         if let Some(block) = &self.pending_block {
             return block.is_satisfied(mem, &self.register_file);
         }
-        self.ring.has_pending()
+        // Pending work may be in the active ring OR in a saved caller ring
+        // further down the IB stack (an exhausted IB still needs `execute_one`
+        // to pop back and resume the primary ring, whose WPTR may have since
+        // advanced).
+        self.ring.has_pending() || self.ib_stack.iter().any(|r| r.has_pending())
     }
 
     /// Execute exactly one PM4 packet. Returns [`ExecOutcome::Idle`] when
@@ -715,10 +772,13 @@ impl GpuSystem {
             width,
             height,
         });
-        self.pending_interrupts.push(PendingInterrupt {
-            source: InterruptSource::Swap,
-            cpu_mask: 0x1,
-        });
+        // iterate-2T: do NOT raise a CP swap-complete interrupt here. Canary's
+        // `VdSwap`/PM4_XE_SWAP path raises no interrupt; swap-complete CP
+        // interrupts come ONLY from in-stream `PM4_INTERRUPT` packets, which
+        // are naturally ordered after D3D has armed the swap-callback slot.
+        // Synthesizing one out of band (as we did pre-2T) delivered a CP
+        // interrupt while the slot still held the `0xBADF00D` placeholder,
+        // tripping the graphics ISR's "Unanticipated CPU_INTERRUPT" assert.
         tracing::info!(
             frame = self.swap_counter,
             fb = format_args!("{frontbuffer_phys:#010x}"),
@@ -730,13 +790,21 @@ impl GpuSystem {
 
     /// Called by `VdInitializeRingBuffer` to give us the primary ring.
     pub fn initialize_ring_buffer(&mut self, base: u32, size_log2: u32) {
-        let size_bytes = 1u32 << size_log2.min(31);
+        // Canary `CommandProcessor::InitializeRingBuffer` (command_processor.cc:
+        // 436): `primary_buffer_size_ = 1 << (size_log2 + 3)` *bytes*. The
+        // `VdInitializeRingBuffer` `r4` argument is log2(size-in-quadwords),
+        // so the byte size is `1 << (size_log2 + 3)` (× 8 bytes/quadword), i.e.
+        // `1 << (size_log2 + 1)` dwords. (Sylpheed passes size_log2=12 →
+        // 32768 bytes / 8192 dwords; the previous `1 << size_log2` undersized
+        // the ring 8× and desynced WPTR wrap math from the guest.)
+        let size_bytes = 1u32 << size_log2.saturating_add(3).min(31);
         // The guest hands us a bare *physical* ring base; project it onto the
         // committed backing window so ring reads hit real PM4 packets (see
         // `physical_to_backing`).
         let base = physical_to_backing(base);
         self.ring.base = base;
         self.ring.size_dwords = size_bytes / 4;
+        self.ring.indirect = false;
         self.ring.read_offset_dwords = 0;
         // `write_offset` is driven by the guest — start at 0 so the ring
         // appears empty until MMIO writes advance it.
@@ -825,6 +893,38 @@ impl GpuSystem {
         }
     }
 
+    /// CP scratch-register memory writeback, mirroring canary's
+    /// `CommandProcessor::HandleSpecialRegisterWrite`
+    /// (`command_processor.cc:545-552`). Every register write runs through
+    /// here; when the target is one of the eight `SCRATCH_REG{n}`
+    /// (`0x0578..=0x057F`) **and** the matching bit in `SCRATCH_UMSK` is set,
+    /// the value is also written (big-endian, as `mem.write_u32` already
+    /// stores) to `SCRATCH_ADDR + n*4` in guest physical memory.
+    ///
+    /// Sylpheed arms its CP swap-complete interrupt callback through this
+    /// path: it programs `SCRATCH_ADDR` to the GPU command-block descriptor
+    /// (`[gfx+10772]`, runtime `0x0b1d5000`), `SCRATCH_UMSK` bit 4, then a
+    /// Type-0 write of the callback PC `0x824ce2b8` into `SCRATCH_REG4`
+    /// (`0x057C`). The writeback lands it at descriptor+16 (`0x4b1d5010`),
+    /// which the graphics ISR (`sub_824BE9A0`) reads via `[[gfx+10772]+16]`
+    /// and `bcctrl`s to fire the swap-complete callback. Without this
+    /// writeback the slot stayed NULL, the ISR skipped the callback, the
+    /// swap counter never advanced, and the title's per-frame manager
+    /// re-fired once then plateaued.
+    fn scratch_register_writeback(&self, mem: &dyn MemoryAccess, index: u32, value: u32) {
+        if !(reg::SCRATCH_REG0..=reg::SCRATCH_REG7).contains(&index) {
+            return;
+        }
+        let scratch_reg = index - reg::SCRATCH_REG0;
+        let umsk = self.register_file.read(reg::SCRATCH_UMSK);
+        if (1u32 << scratch_reg) & umsk == 0 {
+            return;
+        }
+        let scratch_addr = self.register_file.read(reg::SCRATCH_ADDR);
+        let mem_addr = physical_to_backing(scratch_addr.wrapping_add(scratch_reg * 4));
+        mem.write_u32(mem_addr, value);
+    }
+
     fn writeback_read_ptr(&mut self, mem: &dyn MemoryAccess) {
         if self.ring.rptr_writeback_addr != 0 && self.ring.is_initialized() {
             mem.write_u32_fence(
@@ -849,6 +949,7 @@ impl GpuSystem {
             let value = mem.read_u32(dword_addr);
             let target = if write_one { base_index } else { base_index + i };
             self.register_file.write(target, value);
+            self.scratch_register_writeback(mem, target, value);
         }
         tracing::trace!(
             base = format_args!("{base_index:#x}"),
@@ -871,6 +972,8 @@ impl GpuSystem {
         let b = mem.read_u32(b_addr);
         self.register_file.write(reg_index_1, a);
         self.register_file.write(reg_index_2, b);
+        self.scratch_register_writeback(mem, reg_index_1, a);
+        self.scratch_register_writeback(mem, reg_index_2, b);
         tracing::trace!(
             r1 = format_args!("{reg_index_1:#x}"),
             r2 = format_args!("{reg_index_2:#x}"),
@@ -935,6 +1038,10 @@ impl GpuSystem {
                     write_offset_dwords: ib_size, // IB is fully-written at jump time
                     rptr_writeback_addr: 0,
                     rptr_writeback_block_dwords: 0,
+                    // Linear sub-stream: drain [0, ib_size) then pop. Never
+                    // wraps, and `sync_with_mmio`'s CP_RB_WPTR must not touch
+                    // it (canary executes IBs through a separate reader).
+                    indirect: true,
                 };
                 tracing::debug!(
                     ib_ptr = format_args!("{ib_ptr:#010x}"),
@@ -1203,7 +1310,146 @@ impl GpuSystem {
                     "gpu: DRAW_INDX captured"
                 );
                 self.last_draw = Some(ds);
+                let host_vertex_count = processed.host_vertex_count;
                 self.last_primitive = Some(processed);
+
+                // iterate-3O: UI-only per-draw geometry capture. Resolves the
+                // real guest vertex window behind this draw (from the active
+                // VS's vertex-fetch constant) so the host UI can replay the
+                // actual splash geometry instead of synthetic shapes. Entirely
+                // inert in headless/deterministic mode (`frame_captures` is
+                // `None`), so the `--gpu-inline` golden is unaffected.
+                if self.frame_captures.is_some() {
+                    let vs_key = self.active_vs_key.unwrap_or(0);
+                    let ps_key = self.active_ps_key.unwrap_or(0);
+                    let parsed_vs = self
+                        .active_vs_key
+                        .and_then(|k| self.shader_blobs.get(&k))
+                        .map(|b| crate::ucode::parse_shader(&b.dwords));
+                    let (idx_src, idx_size) = match ds.index_source {
+                        crate::draw_state::IndexSource::Dma { index_size, .. } => {
+                            (ds.index_source, index_size)
+                        }
+                        crate::draw_state::IndexSource::Immediate { index_size } => {
+                            (ds.index_source, index_size)
+                        }
+                        crate::draw_state::IndexSource::AutoIndex => {
+                            (ds.index_source, crate::draw_state::IndexSize::Sixteen)
+                        }
+                    };
+                    let cap = crate::draw_capture::build(
+                        self.stats.draws_seen as u32,
+                        ds.primitive,
+                        host_vertex_count,
+                        idx_src,
+                        idx_size,
+                        vs_key,
+                        ps_key,
+                        parsed_vs.as_ref(),
+                        &self.register_file,
+                        mem,
+                    );
+                    if let Some(caps) = self.frame_captures.as_mut() {
+                        // Bound the per-frame list so a runaway frame can't grow
+                        // host memory without limit; keep the most recent.
+                        const MAX_CAPS: usize = 4096;
+                        if caps.len() >= MAX_CAPS {
+                            caps.remove(0);
+                        }
+                        caps.push(cap);
+                    }
+                }
+
+                // P5b: decode the textures the *active pixel shader* actually
+                // samples. Parse the bound PS, collect its `tfetch`
+                // fetch-constant slots, read each 6-dword fetch constant from
+                // the register file, and decode+cache it. `vd_swap` publishes
+                // the result. Empty for flat (no-tfetch) shaders — the
+                // dominant case on Sylpheed's current splash, where this stays
+                // inert until the textured logo draw is reached.
+                self.last_draw_textures.clear();
+                if let Some(ps_key) = self.active_ps_key {
+                    // Collect slots under an immutable borrow of `shader_blobs`,
+                    // then drop it before mutating `texture_cache`.
+                    let slots: Vec<u8> = match self.shader_blobs.get(&ps_key) {
+                        Some(blob) => {
+                            let parsed = crate::ucode::parse_shader(&blob.dwords);
+                            crate::shader_metrics::tfetch_slots(&parsed)
+                        }
+                        None => Vec::new(),
+                    };
+                    for slot in slots {
+                        let mut fetch6 = [0u32; 6];
+                        for (k, w) in fetch6.iter_mut().enumerate() {
+                            *w = self
+                                .register_file
+                                .read(CONST_BASE_FETCH + slot as u32 * 6 + k as u32);
+                        }
+                        let Some(mut key) = crate::texture_cache::decode_fetch_constant(fetch6) else {
+                            continue;
+                        };
+                        // The Xenos texture fetch constant carries a guest
+                        // *physical* base address (`base >> 12`). On the Xbox
+                        // 360 the GPU reads the unified physical memory; the
+                        // CPU writes the (decompressed) texels through its
+                        // cached-physical aperture, which ours backs at the
+                        // committed `0x4000_0000` window. Map the physical
+                        // base onto that backing window so the GPU samples the
+                        // bytes the guest actually wrote — exactly as the
+                        // vertex-fetch path does (`draw_capture.rs`) and as
+                        // canary reads textures through its GPU shared memory
+                        // (= physical). Without this the decode reads the
+                        // low VA `0x0dbee000` (always zero) instead of the
+                        // filled `0x4dbee000`, flattening every disk-asset
+                        // texture (e.g. the publisher logo `E59B2B3D`).
+                        key.base_address = physical_to_backing(key.base_address);
+                        let bi = key.format.block_info();
+                        let span_bytes = (key.pitch_texels as u32)
+                            * (key.height as u32)
+                            * (bi.bytes_per_block as u32)
+                            / (bi.block_w as u32);
+                        let version = span_max_version(mem, key.base_address, span_bytes.max(4));
+                        match self.texture_cache.ensure_cached(key, version, mem) {
+                            Ok(entry) => {
+                                // iterate-3AD: carry the real content `version`
+                                // (from `span_max_version`) so the UI host
+                                // texture cache re-uploads when the guest fills
+                                // more of an evolving atlas (e.g. the 2nd splash
+                                // logo's texels land after the publisher's, in
+                                // the SAME K8888 surface). Previously the UI
+                                // pinned `version_when_uploaded = 1`, so the
+                                // first (partial) upload stuck and later draws
+                                // sampled the not-yet-filled region as black.
+                                self.last_draw_textures
+                                    .push((entry.key, version, entry.bytes.clone()));
+                                metrics::counter!(
+                                    "gpu.texture.decode",
+                                    "fmt" => format!("{:?}", key.format),
+                                )
+                                .increment(1);
+                            }
+                            Err(e) => {
+                                metrics::counter!(
+                                    "gpu.texture.reject",
+                                    "reason" => format!("{e:?}"),
+                                )
+                                .increment(1);
+                            }
+                        }
+                    }
+                }
+
+                // iterate-3T: attach this draw's decoded textures to the just-
+                // captured draw so the UI can bind the real artwork per-draw
+                // (keyed off the active PS's real tfetch slots) instead of a
+                // single last-draw `primary_texture`. UI-only (`frame_captures`
+                // is `None` headless); does not touch the deterministic core.
+                if !self.last_draw_textures.is_empty()
+                    && let Some(caps) = self.frame_captures.as_mut()
+                    && let Some(last) = caps.last_mut()
+                {
+                    last.textures = self.last_draw_textures.clone();
+                }
             }
             pm4::PM4_SET_CONSTANT | pm4::PM4_SET_SHADER_CONSTANTS => {
                 // payload[0] = offset_type — bits[10:0] index, bits[23:16] type
@@ -1483,11 +1729,31 @@ pub mod reg {
     /// `XE_GPU_REG_D1MODE_VBLANK_VLINE_STATUS` (Canary register_table.inc:1126).
     /// Bit 0 = VBLANK_INT_OCCURRED.
     pub const D1MODE_VBLANK_VLINE_STATUS: u32 = 0x1951;
+    /// `XE_GPU_REG_D1MODE_VIEWPORT_SIZE` / `AVIVO_D1MODE_VIEWPORT_SIZE`
+    /// (Canary `register_table.inc:1134`). Packs the active display resolution
+    /// as `(width << 16) | height` with 12-bit fields. The guest's
+    /// swap-complete interrupt callback (`sub_824CE2B8`) divides by the low
+    /// 12 bits (`height`) as a refresh-pacing term, so a 0 read makes its
+    /// `twi` divide-by-zero guard trap and abort the ISR before it clears the
+    /// swap-acknowledge fence. Canary returns the constant below from
+    /// `GraphicsSystem::ReadRegister` (graphics_system.cc:311).
+    pub const D1MODE_VIEWPORT_SIZE: u32 = 0x1961;
     /// `XE_GPU_REG_VGT_EVENT_INITIATOR` — set by EVENT_WRITE.
     pub const VGT_EVENT_INITIATOR: u32 = 0x21F9;
     /// `XE_GPU_REG_COHER_STATUS_HOST` — coherency bits
     /// (Canary `register_table.inc:530`).
     pub const COHER_STATUS_HOST: u32 = 0x0A31;
+    /// `XE_GPU_REG_SCRATCH_UMSK` — bitmask of which `SCRATCH_REG{n}` writes are
+    /// mirrored to memory (Canary `register_table.inc:139`).
+    pub const SCRATCH_UMSK: u32 = 0x01DC;
+    /// `XE_GPU_REG_SCRATCH_ADDR` — base physical address of the scratch
+    /// writeback block (Canary `register_table.inc:141`).
+    pub const SCRATCH_ADDR: u32 = 0x01DD;
+    /// `XE_GPU_REG_SCRATCH_REG0` — first of 8 CP scratch registers
+    /// (`0x0578..=0x057F`, Canary `register_table.inc:331-338`).
+    pub const SCRATCH_REG0: u32 = 0x0578;
+    /// `XE_GPU_REG_SCRATCH_REG7` — last CP scratch register.
+    pub const SCRATCH_REG7: u32 = 0x057F;
 }
 
 /// 32-bit FNV-1a over a u32 seed + a slice of u32s. Used to derive a
@@ -1578,6 +1844,38 @@ mod tests {
         assert_eq!(gpu.register_file.read(0x101), 0xCAFE_BABE);
     }
 
+    #[test]
+    fn scratch_reg_write_mirrors_to_memory_when_umsk_enabled() {
+        // Mirrors Sylpheed's CP swap-callback arming: SCRATCH_ADDR points at a
+        // descriptor, SCRATCH_UMSK enables bit 4, and a Type-0 write of the
+        // callback PC into SCRATCH_REG4 (0x57C) must land at SCRATCH_ADDR + 16.
+        let mut gpu = GpuSystem::new();
+        let mut mem = build_mem();
+        gpu.initialize_ring_buffer(0x4000_0000, 10);
+        // Program SCRATCH_ADDR = 0x4000_1000 (physical-mirror identity), and
+        // SCRATCH_UMSK = bit 4 only (so SCRATCH_REG4 mirrors, REG3 does not).
+        gpu.register_file.write(reg::SCRATCH_ADDR, 0x4000_1000);
+        gpu.register_file.write(reg::SCRATCH_UMSK, 1 << 4);
+        // Type0 write run: base = SCRATCH_REG3 (0x57B), count = 2 → writes
+        // 0x11111111 → SCRATCH_REG3 (UMSK bit 3 clear), 0x824CE2B8 →
+        // SCRATCH_REG4 (UMSK bit 4 set → mirrored to ADDR + 4*4 = +16).
+        const SCRATCH_REG3: u32 = 0x057B;
+        let hdr = (1u32 << 16) | SCRATCH_REG3;
+        mem.write_u32(0x4000_0000, hdr);
+        mem.write_u32(0x4000_0004, 0x1111_1111);
+        mem.write_u32(0x4000_0008, 0x824C_E2B8);
+        gpu.extend_write_ptr(3);
+        assert!(matches!(gpu.execute_one(&mut mem), ExecOutcome::Stepped { .. }));
+        // SCRATCH_REG3 (bit 3 clear) must NOT mirror; SCRATCH_REG4 (bit 4 set)
+        // must mirror to SCRATCH_ADDR + 16.
+        assert_eq!(mem.read_u32(0x4000_1000 + 12), 0, "reg3 must not mirror");
+        assert_eq!(
+            mem.read_u32(0x4000_1000 + 16),
+            0x824C_E2B8,
+            "reg4 must mirror to SCRATCH_ADDR+16"
+        );
+    }
+
     #[test]
     fn wait_reg_mem_blocks_then_unblocks_when_mem_changes() {
         let mut gpu = GpuSystem::new();

crates/xenia-gpu/src/handle.rs

@@ -444,6 +444,23 @@ impl GpuBackend {
         }
     }
 
+    /// Current guest present (`VdSwap`) count. Cheap single-field read used
+    /// by the present-anchored vsync ticker (iterate-3AJ) every scheduler
+    /// round. Inline mode reads the live counter directly; threaded mode
+    /// reads the last-published digest mirror under a brief lock (the
+    /// `--parallel` path uses the wall-clock vsync ticker anyway, so the
+    /// exact freshness here is not load-bearing).
+    pub fn swaps_seen(&self) -> u64 {
+        match self {
+            GpuBackend::Inline(s) => s.stats.swaps_seen,
+            GpuBackend::Threaded(h) => h
+                .digest
+                .lock()
+                .map(|d| d.stats.swaps_seen)
+                .unwrap_or(0),
+        }
+    }
+
     /// Forward [`GpuSystem::has_pending_interrupts`] under inline mode;
     /// under threaded mode peek the `int_rx` channel.
     pub fn has_pending_interrupts(&self) -> bool {

crates/xenia-gpu/src/lib.rs

@@ -12,6 +12,7 @@
 //! [`gpu_system::GpuSystem`].
 
 pub mod command_processor;
+pub mod draw_capture;
 pub mod draw_state;
 pub mod edram;
 pub mod gpu_system;

crates/xenia-gpu/src/mmio_region.rs

@@ -58,6 +58,15 @@ pub fn build_region(mmio: &GpuMmio) -> MmioRegion {
                 reg::D1MODE_VBLANK_VLINE_STATUS => {
                     read_vblank_status.load(Ordering::Relaxed)
                 }
+                // AVIVO_D1MODE_VIEWPORT_SIZE: the active display resolution
+                // (1280x720) packed as `(width << 16) | height`. Canary
+                // serves this constant from `GraphicsSystem::ReadRegister`
+                // (graphics_system.cc:311). The guest swap-complete interrupt
+                // callback divides by the low 12 bits (`height = 0x2D0`); a 0
+                // read trips its `twi` divide-guard and aborts the ISR before
+                // it acknowledges the per-present swap fence — which strands
+                // the present/title loop. Mirror canary exactly.
+                reg::D1MODE_VIEWPORT_SIZE => 0x0500_02D0,
                 _ => {
                     tracing::trace!(
                         reg = format_args!("{reg_index:#x}"),

crates/xenia-gpu/src/primitive.rs

@@ -5,9 +5,8 @@
 //! rectangles) we rewrite indices on the CPU side so the host just sees a
 //! triangle list. Ground truth: `xenia-canary/src/xenia/gpu/primitive_processor.h/cc`.
 //!
-//! P3 scope: only the shapes Sylpheed's UI + early gameplay paths need
-//! (list, strip, fan). Rectangle + quad expansions are stubs logged via
-//! `tracing::warn!` for later.
+//! Scope: list, strip, fan, quad, and rectangle expansions are all handled
+//! (rectangles via CPU triangle-list rewrite — see `expand_rectangles`).
 
 use crate::draw_state::{IndexSize, PrimitiveType};
 
@@ -138,18 +137,43 @@ fn expand_quads(indices: Option<&[u32]>, vertex_count: u32) -> ProcessedPrimitiv
 }
 
 /// Rectangle lists: a Xenos-specific primitive where each group of 3
-/// vertices defines a right-angle rectangle by its three non-repeated
-/// corners (the 4th is derived). The uber-shader doesn't support this yet;
-/// the ucode translator will emulate it as a geometry-stage fake. For P3
-/// we emit an empty draw.
-fn expand_rectangles(_indices: Option<&[u32]>, _vertex_count: u32) -> ProcessedPrimitive {
-    tracing::warn!("gpu: rectangle list primitive not yet implemented (P3 stub)");
-    metrics::counter!("gpu.primitive.rejected", "reason" => "rectangle_list").increment(1);
+/// vertices defines a rectangle; the 4th corner is extrapolated as
+/// `v3 = v0 + v2 - v1` (parallelogram completion). Canary expands this in a
+/// host vertex-shader variant (`kRectangleListAsTriangleStrip`,
+/// `primitive_processor.cc:389-456`): a 4-vertex triangle strip per rect with
+/// the 4th corner synthesized *in the VS* from the host-vertex index.
+///
+/// Our replay pipeline has no host-VS corner synthesis (and the procedural
+/// `vs_main` does not consume `rewritten_indices` yet), so we mirror the
+/// `expand_quads`/`expand_fan` CPU idiom and emit the 3 real vertices of each
+/// rect as one triangle list `(v0,v1,v2)` — the visible lower half of the
+/// rect. This un-rejects the draw and gives a faithful `host_vertex_count`.
+///
+/// TODO: once `vs_main` does real vertex fetch + interpolation, upgrade to the
+/// full quad — 6 indices `[v0,v1,v2, v2,v1,v3]` with a synthesized `v3` corner
+/// — mirroring canary's `kRectangleListAsTriangleStrip`.
+fn expand_rectangles(indices: Option<&[u32]>, vertex_count: u32) -> ProcessedPrimitive {
+    let rect_count = vertex_count / 3;
+    let mut out = Vec::with_capacity(3 * rect_count as usize);
+    let get = |i: u32| -> u32 {
+        match indices {
+            Some(buf) => buf[i as usize],
+            None => i,
+        }
+    };
+    for r in 0..rect_count {
+        let base = r * 3;
+        out.push(get(base));
+        out.push(get(base + 1));
+        out.push(get(base + 2));
+    }
+    let host_vertex_count = out.len() as u32;
+    metrics::counter!("gpu.primitive.expanded", "shape" => "rectangle_list").increment(1);
     ProcessedPrimitive {
         topology: HostTopology::TriangleList,
-        rewritten_indices: Some(Vec::new()),
-        host_vertex_count: 0,
-        rejected: true,
+        rewritten_indices: Some(out),
+        host_vertex_count,
+        rejected: false,
     }
 }
 
@@ -213,6 +237,17 @@ mod tests {
         assert_eq!(idx, vec![0, 1, 2, 0, 2, 3, 4, 5, 6, 4, 6, 7]);
     }
 
+    #[test]
+    fn rectangle_list_expansion() {
+        // 2 rects (6 verts) → one triangle (v0,v1,v2) per rect, not rejected.
+        let p = process(PrimitiveType::RectangleList, 6, None);
+        let idx = p.rewritten_indices.unwrap();
+        assert_eq!(idx, vec![0, 1, 2, 3, 4, 5]);
+        assert_eq!(p.topology, HostTopology::TriangleList);
+        assert_eq!(p.host_vertex_count, 6);
+        assert!(!p.rejected);
+    }
+
     #[test]
     fn widen_u16_indices_big_endian() {
         // 3 indices [1, 2, 0x1234] in BE u16.

crates/xenia-gpu/src/ring_view.rs

@@ -32,6 +32,16 @@ pub struct RingBufferView {
     /// `VdEnableRingBufferRPtrWriteBack`). We always write back eagerly, so
     /// we don't actually use this for scheduling — kept for observability.
     pub rptr_writeback_block_dwords: u32,
+    /// True for an indirect-buffer (`INDIRECT_BUFFER`) view. An IB is a fixed
+    /// *linear* sub-stream, not a circular ring: it is fully written when the
+    /// GPU jumps to it, so the read pointer advances monotonically from `0` to
+    /// `size_dwords` and then the buffer is exhausted (the caller ring is
+    /// popped). It must NOT wrap, and the primary `CP_RB_WPTR` must not be
+    /// applied to it. Mirrors canary `ExecuteIndirectBuffer`, which executes
+    /// the IB through a separate `RingBuffer reader_` and restores the primary
+    /// reader afterward (command_processor.cc). Circular (primary-ring)
+    /// semantics are used when this is `false`.
+    pub indirect: bool,
 }
 
 impl RingBufferView {
@@ -46,7 +56,16 @@ impl RingBufferView {
 
     /// True if there is pending unread data to consume.
     pub fn has_pending(&self) -> bool {
-        self.is_initialized() && self.read_offset_dwords != self.write_offset_dwords
+        if !self.is_initialized() {
+            return false;
+        }
+        if self.indirect {
+            // Linear sub-stream: exhausted once the read pointer reaches the
+            // (fixed) write pointer. Never wraps.
+            self.read_offset_dwords < self.write_offset_dwords
+        } else {
+            self.read_offset_dwords != self.write_offset_dwords
+        }
     }
 
     /// Number of dwords we can consume without wrapping past the write ptr.
@@ -54,7 +73,10 @@ impl RingBufferView {
         if !self.is_initialized() {
             return 0;
         }
-        if self.write_offset_dwords >= self.read_offset_dwords {
+        if self.indirect {
+            self.write_offset_dwords
+                .saturating_sub(self.read_offset_dwords)
+        } else if self.write_offset_dwords >= self.read_offset_dwords {
             self.write_offset_dwords - self.read_offset_dwords
         } else {
             // write has wrapped — we can read up to the end of the ring.
@@ -62,14 +84,20 @@ impl RingBufferView {
         }
     }
 
-    /// Advance the read pointer by `dwords`, wrapping at `size_dwords`.
+    /// Advance the read pointer by `dwords`. Circular rings wrap at
+    /// `size_dwords`; an indirect buffer advances linearly (no wrap) so it
+    /// terminates exactly at its fixed write pointer.
     pub fn advance_read(&mut self, dwords: u32) {
         if self.size_dwords == 0 {
             return;
         }
+        if self.indirect {
+            self.read_offset_dwords = self.read_offset_dwords.saturating_add(dwords);
+        } else {
             self.read_offset_dwords =
                 (self.read_offset_dwords + dwords) % self.size_dwords;
         }
+    }
 
     /// Guest address for the dword at relative offset `i` from the current
     /// read pointer. `None` if uninitialized.
@@ -77,7 +105,11 @@ impl RingBufferView {
         if !self.is_initialized() {
             return None;
         }
-        let off = (self.read_offset_dwords + offset_dwords) % self.size_dwords;
+        let off = if self.indirect {
+            self.read_offset_dwords.saturating_add(offset_dwords)
+        } else {
+            (self.read_offset_dwords + offset_dwords) % self.size_dwords
+        };
         Some(self.base.wrapping_add(off.wrapping_mul(4)))
     }
 }
@@ -120,4 +152,52 @@ mod tests {
         assert_eq!(v.addr_at_offset(1), Some(0x4000_0000));
         assert_eq!(v.addr_at_offset(2), Some(0x4000_0004));
     }
+
+    #[test]
+    fn indirect_buffer_drains_linearly_and_terminates() {
+        // An indirect buffer is a fixed linear sub-stream: read advances from
+        // 0 to `size_dwords` and then is exhausted — it must NOT wrap back to
+        // 0 (which previously caused an infinite re-read of a system command
+        // buffer; iterate-2O). write_offset == size, exactly as the
+        // INDIRECT_BUFFER handler sets it.
+        let mut ib = RingBufferView {
+            base: 0x4adf_5080,
+            size_dwords: 11,
+            read_offset_dwords: 0,
+            write_offset_dwords: 11,
+            rptr_writeback_addr: 0,
+            rptr_writeback_block_dwords: 0,
+            indirect: true,
+        };
+        assert!(ib.has_pending());
+        // Drain the exact packet layout observed for Sylpheed's init IB:
+        // 2 + 3 + 6 dwords = 11.
+        ib.advance_read(2);
+        assert!(ib.has_pending());
+        ib.advance_read(3);
+        assert!(ib.has_pending());
+        ib.advance_read(6); // reaches 11 == write
+        assert_eq!(ib.read_offset_dwords, 11);
+        assert!(
+            !ib.has_pending(),
+            "indirect buffer must terminate at write ptr, not wrap to 0"
+        );
+        // addr_at_offset must not modulo-wrap for an indirect buffer.
+        ib.read_offset_dwords = 9;
+        assert_eq!(ib.addr_at_offset(1), Some(0x4adf_5080 + 10 * 4));
+    }
+
+    #[test]
+    fn indirect_flag_does_not_affect_circular_ring() {
+        // Sanity: a circular (primary) ring still wraps as before.
+        let mut v = RingBufferView::new();
+        v.base = 0x4adc_c000;
+        v.size_dwords = 8192;
+        v.read_offset_dwords = 8190;
+        v.write_offset_dwords = 2;
+        assert!(v.has_pending());
+        v.advance_read(4); // (8190 + 4) % 8192 = 2
+        assert_eq!(v.read_offset_dwords, 2);
+        assert!(!v.has_pending());
+    }
 }

crates/xenia-gpu/src/shader_metrics.rs

@@ -45,8 +45,9 @@ pub fn emit_for(parsed: &ParsedShader, stage: &'static str) {
                         parsed.instructions[base + 1],
                         parsed.instructions[base + 2],
                     ];
-                    // sequence bit layout: 2 bits per triple, hi bit = is-fetch.
-                    let is_fetch = ((sequence >> (i * 2 + 1)) & 1) != 0;
+                    // sequence: 2 bits per instruction — bit[0]=fetch(1)/ALU(0),
+                    // bit[1]=serialize (Xenos `ucode.h:226`).
+                    let is_fetch = ((sequence >> (i * 2)) & 1) != 0;
                     if is_fetch {
                         match decode_fetch(words) {
                             FetchInstruction::Vertex(_) => vfetch_count += 1,
@@ -174,6 +175,50 @@ pub fn emit_for(parsed: &ParsedShader, stage: &'static str) {
     }
 }
 
+/// Collect the unique texture-fetch-constant slot indices a shader samples.
+///
+/// Walks the same exec-clause / sequence-bitmap path as [`emit_for`] but only
+/// extracts `TextureFetch.fetch_const` slots, deduplicated and in first-seen
+/// order. The GPU draw handler uses this to decide which fetch constants to
+/// decode + cache at draw time (keyed off the *active* pixel shader's real
+/// `tfetch` instructions rather than a hardcoded slot).
+pub fn tfetch_slots(parsed: &ParsedShader) -> Vec<u8> {
+    let mut slots: Vec<u8> = Vec::new();
+    for clause in &parsed.cf {
+        if let ControlFlowInstruction::Exec {
+            address,
+            count,
+            sequence,
+            ..
+        } = clause
+        {
+            for i in 0..(*count as usize) {
+                let base = (*address as usize + i) * 3;
+                if base + 2 >= parsed.instructions.len() {
+                    break;
+                }
+                // sequence: 2 bits per instruction — bit[0]=fetch(1)/ALU(0),
+                // bit[1]=serialize (Xenos `ucode.h:226`).
+                let is_fetch = ((sequence >> (i * 2)) & 1) != 0;
+                if !is_fetch {
+                    continue;
+                }
+                let words = [
+                    parsed.instructions[base],
+                    parsed.instructions[base + 1],
+                    parsed.instructions[base + 2],
+                ];
+                if let FetchInstruction::Texture(tf) = decode_fetch(words) {
+                    if !slots.contains(&tf.fetch_const) {
+                        slots.push(tf.fetch_const);
+                    }
+                }
+            }
+        }
+    }
+    slots
+}
+
 fn mark_feature(buf: &mut Vec<&'static str>, name: &'static str) {
     if !buf.contains(&name) {
         buf.push(name);
@@ -298,6 +343,46 @@ mod tests {
         emit_for(&shader, "vs");
     }
 
+    /// `tfetch_slots` should extract the fetch-constant slot of a texture
+    /// fetch (and dedup), and return empty for a flat ALU-only shader.
+    #[test]
+    fn tfetch_slots_extracts_texture_fetch_constants() {
+        // word0: opcode TEXTURE_FETCH (0x01) in low 5 bits, const_index=3 in
+        // bits[24:20] (Xenos `ucode.h:844`) → 0x01 | (3 << 20).
+        let tfetch_w0: u32 = 0x01 | (3u32 << 20);
+        let shader = ParsedShader {
+            cf: vec![
+                ControlFlowInstruction::Exec {
+                    address: 0,
+                    count: 2,
+                    // instruction 0 is a fetch (bit[0] of its 2-bit field set),
+                    // instruction 1 is ALU. is_fetch = (sequence >> (i*2)) & 1.
+                    sequence: 0b00_01,
+                    is_end: false,
+                    predicated: false,
+                    predicate_condition: false,
+                },
+                ControlFlowInstruction::Exit,
+            ],
+            instructions: vec![tfetch_w0, 0, 0, /* ALU triple */ 0, 0, 0],
+        };
+        assert_eq!(tfetch_slots(&shader), vec![3]);
+
+        // Flat shader: no fetch bits → no slots.
+        let flat = ParsedShader {
+            cf: vec![ControlFlowInstruction::Exec {
+                address: 0,
+                count: 1,
+                sequence: 0,
+                is_end: false,
+                predicated: false,
+                predicate_condition: false,
+            }],
+            instructions: vec![0, 0, 0],
+        };
+        assert!(tfetch_slots(&flat).is_empty());
+    }
+
     /// P8: a shader containing `LoopStart` should mark `cf_loop` as used
     /// so the HUD can surface which deferred feature a game triggers.
     #[test]

crates/xenia-gpu/src/shaders/xenos_interp.wgsl

@@ -20,7 +20,15 @@ struct XenosDrawConstants {
     draw_index: u32,
     vertex_count: u32,
     prim_kind: u32,
-    _pad: u32,
+    // iterate-3O: guest dword address that maps to index 0 of `vertex_buffer`.
+    // The CPU uploads a bounded guest-memory window starting at the active
+    // vertex-fetch base; the shader subtracts this base from the absolute
+    // fetch-constant address so it indexes the uploaded window. 0 means "no
+    // real vertex window" (procedural fallback path).
+    vertex_base_dwords: u32,
+    // iterate-3S: guest viewport → host NDC XY transform (Y pre-flipped).
+    ndc_scale: vec2<f32>,
+    ndc_offset: vec2<f32>,
 };
 
 struct XenosConstants {
@@ -56,6 +64,7 @@ const CF_KIND_LOOP_END:    u32 = 5u;
 const CF_KIND_COND_JMP:    u32 = 6u;
 const CF_KIND_COND_CALL:   u32 = 7u;
 const CF_KIND_RETURN:      u32 = 8u;
+const CF_KIND_NOP:         u32 = 9u;
 const CF_KIND_UNKNOWN:     u32 = 15u;
 
 // ── Alloc-kind codes (mirrors `xenia_gpu::ucode::cf_alloc_kind`). ──────
@@ -628,8 +637,8 @@ const VFMT_32_32_32_FLOAT:  u32 = 57u;
 // layout in `ucode.h:690`):
 //   w0 [4:0]   opcode
 //   w0 [10:5]  src_reg[5:0]
-//   w0 [17:11] dst_reg[6:0] + must-be-one
-//   w0 [21:17] const_index[4:0], [23:22] const_index_sel[1:0]
+//   w0 [17:12] dst_reg[5:0]
+//   w0 [24:20] const_index[4:0], [26:25] const_index_sel[1:0]
 //   w1 [21:16] format[5:0]
 //   w2 [7:0]   stride (in dwords)
 //   w2 [30:8]  offset (signed, in dwords)
@@ -641,9 +650,9 @@ fn interpret_vertex_fetch(t: u32) {
     let w0 = vs_instr_dword(t, 0u);
     let w1 = vs_instr_dword(t, 1u);
     let w2 = vs_instr_dword(t, 2u);
-    let fetch_const = (w0 >> 5u) & 0x1Fu;
-    let dst_reg = (w0 >> 10u) & 0x7Fu;
-    let src_reg = (w0 >> 17u) & 0x7Fu;
+    let fetch_const = (w0 >> 20u) & 0x1Fu;
+    let dst_reg = (w0 >> 12u) & 0x3Fu;
+    let src_reg = (w0 >> 5u) & 0x3Fu;
     let format  = (w1 >> 16u) & 0x3Fu;
     let stride  = w2 & 0xFFu;
 
@@ -651,7 +660,20 @@ fn interpret_vertex_fetch(t: u32) {
     // dword 1 carries (endian[1:0], size[25:2]).
     let fc0 = xenos_consts.fetch[fetch_const * 2u + 0u];
     let fc1 = xenos_consts.fetch[fetch_const * 2u + 1u];
-    let base_dwords = (fc0 & 0xFFFFFFFCu) >> 2u;
+    // iterate-3O: the fetch constant holds an *absolute* guest dword address.
+    // The CPU uploaded a window of guest memory starting at
+    // `draw_ctx.vertex_base_dwords`, so rebase the absolute address into that
+    // window. When no real window was published (`vertex_base_dwords == 0`)
+    // keep the absolute value (the `addr < n` guards below then skip the read
+    // and the procedural fallback position is used).
+    // GPUBUG-108 (iterate-3S): the captured window begins exactly at the fetch
+    // base, so index from 0 (vertex i at i*stride). The uniform `fetch[]` holds
+    // the last-published per-frame constant, not this draw's — recomputing
+    // `abs_base` from it produced a stale out-of-window address (the splash
+    // collapsed to one pixel). Only consult the uniform for the no-window
+    // synthetic fallback.
+    let abs_base = (fc0 & 0xFFFFFFFCu) >> 2u;
+    let base_dwords = select(abs_base, 0u, draw_ctx.vertex_base_dwords != 0u);
     // GPUBUG-102: per-format endian byte-swap. Xbox 360 vertex data is
     // big-endian; the host is little-endian. Pre-fix every dword was
     // bitcast as-is — vertex positions were byte-reversed garbage.
@@ -773,20 +795,20 @@ fn interpret_texture_fetch(t: u32, is_vertex: bool) {
     } else {
         w0 = ps_instr_dword(t, 0u);
     }
-    let dst_reg = (w0 >> 10u) & 0x7Fu;
-    let src_reg = (w0 >> 17u) & 0x7Fu;
-    let uv = registers[src_reg & 0x7Fu].xy;
+    let dst_reg = (w0 >> 12u) & 0x3Fu;
+    let src_reg = (w0 >> 5u) & 0x3Fu;
+    let uv = registers[src_reg & 0x3Fu].xy;
     let sample = textureSampleLevel(xenos_tex, xenos_samp, uv, 0.0);
-    registers[dst_reg & 0x7Fu] = sample;
+    registers[dst_reg & 0x3Fu] = sample;
 }
 
 // Walk an Exec clause's instruction triples.
-//   sequence: 2-bit-per-triple bitmap. Bit 0 of a pair = serialize flag
-//             (we ignore in MVP); bit 1 = is-fetch.
+//   sequence: 2-bit-per-instruction bitmap. Bit 0 of a pair = fetch(1)/ALU(0);
+//             bit 1 = serialize (ignored). (Xenos `ucode.h:226`.)
 fn exec_vs(address: u32, count: u32, sequence: u32) {
     for (var i: u32 = 0u; i < count; i = i + 1u) {
         let t = address + i;
-        let is_fetch = ((sequence >> (i * 2u + 1u)) & 1u) != 0u;
+        let is_fetch = ((sequence >> (i * 2u)) & 1u) != 0u;
         if is_fetch {
             let opcode = vs_instr_dword(t, 0u) & 0x1Fu;
             // 0x00 = vertex fetch, 0x01 = texture fetch.
@@ -803,7 +825,7 @@ fn exec_vs(address: u32, count: u32, sequence: u32) {
 fn exec_ps(address: u32, count: u32, sequence: u32) {
     for (var i: u32 = 0u; i < count; i = i + 1u) {
         let t = address + i;
-        let is_fetch = ((sequence >> (i * 2u + 1u)) & 1u) != 0u;
+        let is_fetch = ((sequence >> (i * 2u)) & 1u) != 0u;
         if is_fetch {
             interpret_texture_fetch(t, false);
         } else {
@@ -871,7 +893,13 @@ fn vs_main(@builtin(vertex_index) vidx: u32) -> VsOut {
     // Use registers[OPOS_REG] as position; the procedural fallback above
     // seeded it so an un-interpreted shader still draws a recognisable
     // circle.
-    out.position = vec4<f32>(registers[OPOS_REG].xyz, registers[OPOS_REG].w);
+    var opos = vec4<f32>(registers[OPOS_REG].xyz, registers[OPOS_REG].w);
+    // iterate-3S: guest VS position → host clip space (see translator.rs). When
+    // the transform is unset (procedural fallback) pass through unchanged.
+    if (draw_ctx.ndc_scale.x != 0.0 || draw_ctx.ndc_scale.y != 0.0) {
+        opos = vec4<f32>(opos.xy * draw_ctx.ndc_scale + draw_ctx.ndc_offset * opos.w, opos.z, opos.w);
+    }
+    out.position = opos;
     out.color = vec4<f32>(registers[OCOLOR_REG].rgb + vec3<f32>(vb_live), registers[OCOLOR_REG].a);
     return out;
 }
@@ -962,6 +990,9 @@ fn walk_cf_vs() {
                 // No call stack — mark and continue.
                 reject_mask |= REJECT_CF_CALL;
             }
+            case CF_KIND_NOP: {
+                // kNop padding / kMarkVsFetchDone hint — no-op, just advance.
+            }
             default: { reject_mask |= REJECT_CF_JUMP; }
         }
         if stop { break; }

crates/xenia-gpu/src/translator.rs

@@ -94,7 +94,9 @@ struct XenosDrawConstants {
     draw_index: u32,
     vertex_count: u32,
     prim_kind: u32,
-    _pad: u32,
+    vertex_base_dwords: u32,
+    ndc_scale: vec2<f32>,
+    ndc_offset: vec2<f32>,
 };
 
 struct XenosConstants {
@@ -113,9 +115,21 @@ struct XenosConstants {
 @group(1) @binding(0) var xenos_tex  : texture_2d<f32>;
 @group(1) @binding(1) var xenos_samp : sampler;
 
+// iterate-3T: real interpolator passthrough. The Xenos VS exports up to 16
+// interpolators (export index 0..15); the PS reads interpolator i from its
+// general register r[i]. We carry 8 interpolator vec4s (covers Sylpheed's
+// splash: r0=color, r1=texcoord). `color` retained as an alias of interp0 so
+// older single-color paths keep working.
 struct VsOut {
     @builtin(position) position: vec4<f32>,
-    @location(0) color: vec4<f32>,
+    @location(0) interp0: vec4<f32>,
+    @location(1) interp1: vec4<f32>,
+    @location(2) interp2: vec4<f32>,
+    @location(3) interp3: vec4<f32>,
+    @location(4) interp4: vec4<f32>,
+    @location(5) interp5: vec4<f32>,
+    @location(6) interp6: vec4<f32>,
+    @location(7) interp7: vec4<f32>,
 };
 
 struct FsOut {
@@ -154,6 +168,14 @@ struct EmitCtx {
     stage: Stage,
     out: String,
     indent: usize,
+    /// GPUBUG-114: dword stride of the most recent *full* vfetch, keyed by
+    /// fetch-const register offset. A vfetch_mini carries stride=0 and reuses
+    /// the address + stride of the preceding full vfetch of the same stream
+    /// (canary ucode.h:733). Without this a mini color attribute indexes by its
+    /// tight dword count instead of the real vertex stride → reads the wrong
+    /// vertex's data (Sylpheed's background fill `0x36660986` read garbage →
+    /// white instead of the intended color).
+    last_full_stride: std::collections::HashMap<u32, u32>,
 }
 
 impl EmitCtx {
@@ -162,6 +184,7 @@ impl EmitCtx {
             stage,
             out: String::with_capacity(2048),
             indent: 0,
+            last_full_stride: std::collections::HashMap::new(),
         }
     }
 
@@ -198,19 +221,74 @@ impl EmitCtx {
         self.push("var ps: f32 = 0.0;");
         match self.stage {
             Stage::Vertex => {
+                // iterate-3T: host→guest vertex-index remap for primitives the
+                // replay draws non-indexed as a flat triangle list. wgpu has no
+                // QuadList/RectangleList topology, so the host issues 6 vertices
+                // per quad/rect and we map them back to the guest's 4/3 source
+                // vertices here (mirrors `primitive.rs` index rewrite, but in the
+                // VS since the replay path is non-indexed):
+                //   QuadList(13):  6 host verts → guest [0,1,2, 0,2,3]
+                //   RectangleList(8): drawn as one triangle [0,1,2] (the 4th
+                //     corner needs cross-vertex synthesis — TODO), so host
+                //     indices >=3 fold onto the existing triangle.
+                // Other prims pass through unchanged.
+                self.push("var gvidx: u32 = vidx;");
+                self.push("if (draw_ctx.prim_kind == 13u) {");
+                self.indent += 1;
+                self.push("let q = vidx % 6u; let qbase = (vidx / 6u) * 4u;");
+                self.push("var lut = array<u32, 6>(0u, 1u, 2u, 0u, 2u, 3u);");
+                self.push("gvidx = qbase + lut[q];");
+                self.indent -= 1;
+                self.push("} else if (draw_ctx.prim_kind == 8u) {");
+                self.indent += 1;
+                self.push("let t = vidx % 3u; let rbase = (vidx / 3u) * 3u;");
+                self.push("gvidx = rbase + t;");
+                self.indent -= 1;
+                self.push("}");
                 // Seed r0 with vertex index for simple shaders that read it.
-                self.push("r[0] = vec4<f32>(f32(vidx), 0.0, 0.0, 1.0);");
-                // Synthetic export slots — match the interpreter's layout so
-                // the fallback path and translator path produce the same
-                // visual output on shaders both support.
+                self.push("r[0] = vec4<f32>(f32(gvidx), 0.0, 0.0, 1.0);");
+                // iterate-3T: real export model. Xenos export index 62 = oPos;
+                // indices 0..15 = interpolators. We hold position + 8
+                // interpolator vec4s; `emit_export` writes the right slot keyed
+                // on the export index.
+                //
+                // iterate-3AE (WHITE-TRIANGLE ROOT): interpolators a VS does NOT
+                // export must default to ZERO, not white. The old `ointerp[0] =
+                // (1,1,1,1)` was an iterate-3T debug convenience ("so a VS that
+                // only exports position still yields a visible non-zero color")
+                // — but it is a FAKE: it injects white that no guest value backs.
+                // The transition/background draws use the position-only VS
+                // `0xd4c14f46` (one vfetch → oPos; it exports NO color) paired
+                // with PS `0xed732b5a` (`ocolor0 = interp0`). With the white
+                // seed, interp0 stayed (1,1,1,1) → the fullscreen fill rendered
+                // OPAQUE WHITE (the diagonal half-triangle artifact that flashed
+                // before each splash logo and persisted across the dev-logo
+                // transition). Canary shows a black background there because the
+                // un-exported interpolator carries no white. Default to
+                // (0,0,0,0): a position-only VS now contributes nothing visible
+                // under its real (opaque or premultiplied) blend, matching
+                // canary, while every VS that really exports interp0 (the logo
+                // `0x03b7b020`, the `0x36660986` color fill) overwrites this seed
+                // and is unaffected. RGB=0 → black fill; A=0 → premultiplied
+                // overlays stay transparent.
                 self.push("var opos: vec4<f32> = vec4<f32>(0.0, 0.0, 0.0, 1.0);");
-                self.push("var ocolor: vec4<f32> = vec4<f32>(1.0, 1.0, 1.0, 1.0);");
+                self.push("var ointerp: array<vec4<f32>, 8>;");
+                self.push("for (var i = 0u; i < 8u; i = i + 1u) { ointerp[i] = vec4<f32>(0.0, 0.0, 0.0, 0.0); }");
             }
             Stage::Pixel => {
-                // Seed r0.xy with interpolated color lane so trivial shaders
-                // that read r0 still produce something.
-                self.push("r[0] = in.color;");
-                self.push("var ocolor0: vec4<f32> = in.color;");
+                // iterate-3T: the PS reads interpolator i from general register
+                // r[i] (Xenos PS input GPR mapping). Seed r0..r7 from the VS's
+                // interpolators so e.g. the logo PS's texcoord (r1) and color
+                // (r0) arrive correctly; tfetch then samples at the real UV.
+                self.push("r[0] = in.interp0;");
+                self.push("r[1] = in.interp1;");
+                self.push("r[2] = in.interp2;");
+                self.push("r[3] = in.interp3;");
+                self.push("r[4] = in.interp4;");
+                self.push("r[5] = in.interp5;");
+                self.push("r[6] = in.interp6;");
+                self.push("r[7] = in.interp7;");
+                self.push("var ocolor0: vec4<f32> = in.interp0;");
             }
         }
 
@@ -237,6 +315,10 @@ impl EmitCtx {
                     current_alloc = *kind;
                 }
                 ControlFlowInstruction::Exit => break,
+                // Non-executing CF clauses: padding (`kNop`) and the
+                // vertex-fetch-done hint (`kMarkVsFetchDone`). Skip them.
+                ControlFlowInstruction::Nop
+                | ControlFlowInstruction::MarkVsFetchDone => {}
                 ControlFlowInstruction::LoopStart { .. }
                 | ControlFlowInstruction::LoopEnd { .. } => return Err(reject::CF_LOOP),
                 ControlFlowInstruction::CondJmp { .. } => return Err(reject::CF_COND),
@@ -250,13 +332,41 @@ impl EmitCtx {
         match self.stage {
             Stage::Vertex => {
                 self.push("var out: VsOut;");
+                // iterate-3S: guest VS position → host clip space. The guest
+                // emits either clip-space or (screen-space, clip disabled)
+                // render-target-pixel coords; `ndc_scale`/`ndc_offset` (from
+                // canary's GetHostViewportInfo, computed CPU-side per draw)
+                // rescale XY into wgpu clip space with Y already flipped. When
+                // the transform is unset (all-zero scale, procedural fallback)
+                // pass the position through unchanged.
+                self.push("if (draw_ctx.ndc_scale.x != 0.0 || draw_ctx.ndc_scale.y != 0.0) {");
+                self.indent += 1;
+                self.push("opos = vec4<f32>(opos.xy * draw_ctx.ndc_scale + draw_ctx.ndc_offset * opos.w, opos.z, opos.w);");
+                self.indent -= 1;
+                self.push("}");
                 self.push("out.position = opos;");
-                self.push("out.color = ocolor;");
+                self.push("out.interp0 = ointerp[0];");
+                self.push("out.interp1 = ointerp[1];");
+                self.push("out.interp2 = ointerp[2];");
+                self.push("out.interp3 = ointerp[3];");
+                self.push("out.interp4 = ointerp[4];");
+                self.push("out.interp5 = ointerp[5];");
+                self.push("out.interp6 = ointerp[6];");
+                self.push("out.interp7 = ointerp[7];");
                 self.push("return out;");
             }
             Stage::Pixel => {
                 self.push("var out: FsOut;");
-                self.push("out.color0 = ocolor0;");
+                // GPUBUG-115: saturate the color export to [0,1], flushing NaN
+                // to 0 — exactly what canary does before writing a UNORM render
+                // target (spirv_shader_translator.cc:3607 "Saturate, flushing
+                // NaN to 0"). The Xenos RB clamps PS output for UNORM targets;
+                // without this an out-of-range guest color (Sylpheed's
+                // background fill exports a huge negative float `-32896.5` as a
+                // fullscreen-clear value) writes garbage/NaN to the sRGB target
+                // → renders white instead of the clamped black canary shows.
+                // `clamp(x,0,1)` returns 0 for NaN under WGSL's clamp semantics.
+                self.push("out.color0 = clamp(ocolor0, vec4<f32>(0.0), vec4<f32>(1.0));");
                 self.push("return out;");
             }
         }
@@ -284,7 +394,9 @@ impl EmitCtx {
                 parsed.instructions[base + 1],
                 parsed.instructions[base + 2],
             ];
-            let is_fetch = ((sequence >> (i * 2 + 1)) & 1) != 0;
+            // sequence: 2 bits per instruction — bit[0]=fetch(1)/ALU(0),
+            // bit[1]=serialize (Xenos `ucode.h:226`).
+            let is_fetch = ((sequence >> (i * 2)) & 1) != 0;
             if is_fetch {
                 match decode_fetch(words) {
                     FetchInstruction::Vertex(vf) => self.emit_vfetch(&vf)?,
@@ -378,53 +490,185 @@ impl EmitCtx {
     }
 
     fn emit_export(&mut self, dst_reg: u8, alloc: AllocKind, expr: &str, mask: u8) {
-        // Xenos's export "register" indexing within an alloc range is
-        // normally (alloc_base + offset). Since our CF stream doesn't
-        // carry per-export slot offsets cleanly, use `alloc` to pick the
-        // target.
-        let lhs = match (self.stage, alloc) {
-            (Stage::Vertex, AllocKind::Position) => "opos",
-            (Stage::Vertex, AllocKind::Interpolators) => "ocolor",
-            (Stage::Vertex, AllocKind::Colors) => "ocolor",
-            (Stage::Vertex, _) => "ocolor", // fall through — any other alloc
-            (Stage::Pixel, AllocKind::Colors) => "ocolor0",
-            (Stage::Pixel, _) => "ocolor0",
+        // iterate-3T: real Xenos export-index model (replaces the `AllocKind`
+        // heuristic, which collapsed every VS export to a single color slot and
+        // dropped the texcoord interpolator → tfetch sampled (0,0) → flat).
+        // When `export_data` is set the 6-bit vector_dest IS the export index:
+        //   VS: 62 = oPos, 63 = oPointSize/edge (ignored), 0..15 = interpolators.
+        //   PS: 0..3 = color render targets (we honor RT0).
+        let _ = alloc;
+        match self.stage {
+            Stage::Vertex => {
+                let lhs = if dst_reg == 62 {
+                    "opos".to_string()
+                } else if dst_reg <= 15 {
+                    // Clamp to the 8 interpolator slots we carry; higher slots
+                    // are unused by Sylpheed's splash.
+                    let i = (dst_reg as usize).min(7);
+                    format!("ointerp[{i}u]")
+                } else {
+                    // oPointSize (63) / unknown export slot — discard.
+                    return;
                 };
-        let _ = dst_reg; // per-slot export indexing reserved for a richer v2
-        self.emit_masked_write(lhs, expr, mask);
+                self.emit_masked_write(&lhs, expr, mask);
+            }
+            Stage::Pixel => {
+                // Only RT0 (export index 0) is wired to the single host target.
+                if dst_reg == 0 {
+                    self.emit_masked_write("ocolor0", expr, mask);
+                }
+            }
+        }
     }
 
     fn emit_vfetch(&mut self, vf: &crate::ucode::fetch::VertexFetch) -> Result<(), &'static str> {
-        // v1: treat all vertex fetches as R32G32B32A32_FLOAT, stride = 4
-        // dwords. Matches the interpreter's MVP semantics; unlocks more
-        // formats alongside the CPU texture cache's format expansion.
+        // GPUBUG-107 (iterate-3S): decode the vertex FORMAT + dword STRIDE from
+        // the vfetch instruction instead of hardcoding R32G32B32A32 (4 floats,
+        // stride 4). Sylpheed's splash quads are `k_32_32_FLOAT` (2 floats,
+        // stride 2); over-reading them put the next vertex's X into .w → a
+        // negative W → the whole rectangle clipped behind the camera. We cover
+        // the float vertex formats (the UI / screen-space draws); other formats
+        // reject to the interpreter.
         //
-        // GPUBUG-102: the fetch constant (xe_gpu_vertex_fetch_t,
-        // xenos.h:1158-1172) holds the endian field in dword_1's low
-        // 2 bits. Vertex data on Xbox 360 is big-endian; the host is
-        // little-endian. Pre-fix, every dword was bitcast as-is →
-        // vertex positions were byte-reversed garbage and any draw
-        // that did reach the host produced clipped / NaN positions.
-        let fetch_const = (vf.raw[0] >> 5) & 0x1F;
+        // GPUBUG-102: the fetch constant holds the endian field in dword_1's
+        // low 2 bits; Xbox 360 vertex data is big-endian, so `gpu_swap` undoes
+        // it per component.
+        // (comps, dwords_read) per format. Float formats are 1 dword/component;
+        // iterate-3T adds the packed-16 `k_16_16` (format 6) used for the logo
+        // UV interpolator — 2 components packed into ONE dword.
+        #[derive(PartialEq)]
+        enum Pack {
+            Float,    // N f32 lanes, N dwords
+            Norm16x2, // 2× u16 normalized into [0,1], 1 dword (k_16_16)
+            Norm8x4,  // 4× u8 normalized into [0,1], 1 dword (k_8_8_8_8)
+        }
+        let (comps, dwords_read, pack): (u32, u32, Pack) = match vf.format {
+            36 => (1, 1, Pack::Float),   // k_32_FLOAT
+            37 => (2, 2, Pack::Float),   // k_32_32_FLOAT
+            57 => (3, 3, Pack::Float),   // k_32_32_32_FLOAT
+            38 => (4, 4, Pack::Float),   // k_32_32_32_32_FLOAT
+            6 => (4, 1, Pack::Norm8x4),  // k_8_8_8_8 (packed RGBA8 — GPUBUG-112)
+            25 => (2, 1, Pack::Norm16x2), // k_16_16
+            _ => return Err(reject::VFETCH_FMT),
+        };
+        // iterate-3X (GPUBUG-110): index the fetch-constant region by the full
+        // `const_index*3 + const_index_sel` mapping (canary `ucode.h:700`),
+        // packed as `const_index*6 + sel*2` dwords. The previous expression
+        // `(vf.raw[0] >> 5) & 0x1F` read the *src_reg* bits, not the const
+        // index — wrong for the endian term and the no-window fallback base.
+        let const_off = vf.const_reg_offset();
+        // GPUBUG-114: a full vfetch carries the real vertex dword stride; a
+        // vfetch_mini reuses the address + stride of the preceding full vfetch
+        // of the same stream (canary ucode.h:733). Track the last full stride
+        // per fetch-const and inherit it for mini-fetches (stride field == 0).
+        let stride = if vf.is_mini_fetch || vf.stride == 0 {
+            *self
+                .last_full_stride
+                .get(&const_off)
+                .unwrap_or(&dwords_read)
+        } else {
+            self.last_full_stride.insert(const_off, vf.stride as u32);
+            vf.stride as u32
+        };
+        // iterate-3T: per-attribute dword offset within the vertex (vfetches
+        // sharing one fetch constant read different attributes).
+        let attr_off = vf.offset;
         let src_reg = vf.src_register & 0x7F;
         let dst_reg = vf.dest_register & 0x7F;
+        // is_signed selects [-1,1] vs [0,1] for normalized integer formats.
+        let signed = vf.is_signed;
+        // Build the per-component reads; unread lanes default to 0/0/0/1 so an
+        // XY-only position keeps W=1 (and Z=0).
+        let lane = |i: u32| -> String {
+            match pack {
+                Pack::Float => {
+                    if i < comps {
+                        format!("bitcast<f32>(gpu_swap(vertex_buffer[addr + {i}u], endian))")
+                    } else if i == 3 {
+                        "1.0".to_string()
+                    } else {
+                        "0.0".to_string()
+                    }
+                }
+                Pack::Norm16x2 => {
+                    // One dword holds [u16 lo | u16 hi] after the endian swap.
+                    // Component 0 = low halfword, component 1 = high halfword.
+                    if i == 0 {
+                        if signed {
+                            "(max(f32(i32(w16 << 16u) >> 16u) / 32767.0, -1.0))".to_string()
+                        } else {
+                            "(f32(w16 & 0xFFFFu) / 65535.0)".to_string()
+                        }
+                    } else if i == 1 {
+                        if signed {
+                            "(max(f32(i32(w16) >> 16u) / 32767.0, -1.0))".to_string()
+                        } else {
+                            "(f32(w16 >> 16u) / 65535.0)".to_string()
+                        }
+                    } else if i == 3 {
+                        "1.0".to_string()
+                    } else {
+                        "0.0".to_string()
+                    }
+                }
+                Pack::Norm8x4 => {
+                    // One dword holds 4× u8 (canary spirv_shader_translator_fetch
+                    // k_8_8_8_8: comp0@bit0, comp1@bit8, comp2@bit16, comp3@bit24)
+                    // after the endian swap. All four channels present → normalize
+                    // to [0,1]. GPUBUG-112: this is the logo/background vertex
+                    // COLOR (RGBA8), previously misdecoded as k_16_16 (2 chans,
+                    // B forced 0) → white texture × (R,G,0) = yellow.
+                    let sh = i * 8;
+                    if signed {
+                        format!(
+                            "(max(f32(i32(w16 << {l}u) >> 24u) / 127.0, -1.0))",
+                            l = 24 - sh
+                        )
+                    } else {
+                        format!("(f32((w16 >> {sh}u) & 0xFFu) / 255.0)")
+                    }
+                }
+            }
+        };
+        let read_bound = dwords_read - 1;
+        // GPUBUG-108 (iterate-3S): for the captured-geometry path the CPU
+        // uploads a vertex window that begins EXACTLY at the fetch base, so the
+        // base within `vertex_buffer` is 0 and vertex i sits at `i * stride`.
+        // The previous `abs_base - vertex_base_dwords` rebase recomputed the
+        // base from `xenos_consts.fetch[]`, but that uniform carries the
+        // *last-published* (per-frame) fetch constant, not this draw's — for
+        // the splash it was stale (0x8a000002 vs the real 0x0adf… base), so the
+        // rebase produced a huge out-of-window address, the bounds guard
+        // failed, and every vertex kept its seed (vertex_index, 0, 0, 1) →
+        // every quad collapsed to ~one pixel at the origin. Index from 0 when a
+        // real window is present (`vertex_base_dwords != 0`); only the
+        // synthetic/no-window fallback consults the uniform fetch constant.
+        let endian_term = format!("xenos_consts.fetch[{}u] & 0x3u", const_off + 1);
+        // For packed formats (k_16_16, k_8_8_8_8) we read one dword into `w16`
+        // (post endian-swap) and the `lane()` exprs above unpack the channels.
+        let w16_decl = if pack == Pack::Norm16x2 || pack == Pack::Norm8x4 {
+            "let w16 = gpu_swap(vertex_buffer[addr], endian); "
+        } else {
+            ""
+        };
         self.push(&format!(
-            "{{ let fc0 = xenos_consts.fetch[{fc0_idx}u]; \
-             let fc1 = xenos_consts.fetch[{fc1_idx}u]; \
-             let endian = fc1 & 0x3u; \
-             let base = (fc0 & 0xFFFFFFFCu) >> 2u; \
+            "{{ let endian = {endian_term}; \
              let vidx = u32(r[{src_reg}u].x); \
-             let addr = base + vidx * 4u; \
+             var base = 0u; \
+             if (draw_ctx.vertex_base_dwords == 0u) {{ \
+                 base = (xenos_consts.fetch[{fc0_idx}u] & 0xFFFFFFFCu) >> 2u; \
+             }} \
+             let addr = base + vidx * {stride}u + {attr_off}u; \
              let n = arrayLength(&vertex_buffer); \
-             if (addr + 3u < n) {{ \
-                 r[{dst_reg}u] = vec4<f32>( \
-                     bitcast<f32>(gpu_swap(vertex_buffer[addr + 0u], endian)), \
-                     bitcast<f32>(gpu_swap(vertex_buffer[addr + 1u], endian)), \
-                     bitcast<f32>(gpu_swap(vertex_buffer[addr + 2u], endian)), \
-                     bitcast<f32>(gpu_swap(vertex_buffer[addr + 3u], endian))); \
+             if (addr + {read_bound}u < n) {{ \
+                 {w16_decl}\
+                 r[{dst_reg}u] = vec4<f32>({l0}, {l1}, {l2}, {l3}); \
              }} }}",
-            fc0_idx = fetch_const * 2,
-            fc1_idx = fetch_const * 2 + 1,
+            fc0_idx = const_off,
+            l0 = lane(0),
+            l1 = lane(1),
+            l2 = lane(2),
+            l3 = lane(3),
         ));
         Ok(())
     }
@@ -477,6 +721,22 @@ fn src_operand(src_byte: u8, is_temp: bool, swizzle: u8, negate: bool) -> String
 }
 
 fn vector_expr(op: u8, a: &str, b: &str, c: &str) -> Option<String> {
+    // Semantics mirror the runtime interpreter's `exec_vector_op`
+    // (`shaders/xenos_interp.wgsl`), which in turn mirrors canary's
+    // `AluVectorOpcode` (ucode.h:1001+). Side-effecting ops (kill*, setp_push)
+    // need per-invocation state the AOT emitter doesn't track yet → still
+    // `None` (interpreter fallback).
+    let cmp4 = |op: &str| {
+        format!(
+            "vec4<f32>(select(0.0,1.0,{a}.x{op}{b}.x), select(0.0,1.0,{a}.y{op}{b}.y), select(0.0,1.0,{a}.z{op}{b}.z), select(0.0,1.0,{a}.w{op}{b}.w))"
+        )
+    };
+    // CND* : per-lane select(c, b, a <cmp> 0).
+    let cnd4 = |op: &str| {
+        format!(
+            "vec4<f32>(select({c}.x,{b}.x,{a}.x{op}0.0), select({c}.y,{b}.y,{a}.y{op}0.0), select({c}.z,{b}.z,{a}.z{op}0.0), select({c}.w,{b}.w,{a}.w{op}0.0))"
+        )
+    };
     let s = match op {
         vop::ADD => format!("({a} + {b})"),
         vop::MUL => format!("({a} * {b})"),
@@ -485,37 +745,63 @@ fn vector_expr(op: u8, a: &str, b: &str, c: &str) -> Option<String> {
         vop::MAD => format!("({a} * {b} + {c})"),
         vop::DOT4 => format!("vec4<f32>(dot({a}, {b}))"),
         vop::DOT3 => format!("vec4<f32>(dot({a}.xyz, {b}.xyz))"),
-        vop::DOT2_ADD => format!(
-            "vec4<f32>({a}.x * {b}.x + {a}.y * {b}.y + {c}.x)"
-        ),
-        vop::SEQ => format!(
-            "vec4<f32>(select(0.0,1.0,{a}.x=={b}.x), select(0.0,1.0,{a}.y=={b}.y), select(0.0,1.0,{a}.z=={b}.z), select(0.0,1.0,{a}.w=={b}.w))"
-        ),
-        vop::SGT => format!(
-            "vec4<f32>(select(0.0,1.0,{a}.x>{b}.x), select(0.0,1.0,{a}.y>{b}.y), select(0.0,1.0,{a}.z>{b}.z), select(0.0,1.0,{a}.w>{b}.w))"
-        ),
-        vop::SGE => format!(
-            "vec4<f32>(select(0.0,1.0,{a}.x>={b}.x), select(0.0,1.0,{a}.y>={b}.y), select(0.0,1.0,{a}.z>={b}.z), select(0.0,1.0,{a}.w>={b}.w))"
-        ),
-        vop::SNE => format!(
-            "vec4<f32>(select(0.0,1.0,{a}.x!={b}.x), select(0.0,1.0,{a}.y!={b}.y), select(0.0,1.0,{a}.z!={b}.z), select(0.0,1.0,{a}.w!={b}.w))"
-        ),
+        vop::DOT2_ADD => format!("vec4<f32>({a}.x * {b}.x + {a}.y * {b}.y + {c}.x)"),
+        vop::SEQ => cmp4("=="),
+        vop::SGT => cmp4(">"),
+        vop::SGE => cmp4(">="),
+        vop::SNE => cmp4("!="),
+        vop::CND_EQ => cnd4("=="),
+        vop::CND_GE => cnd4(">="),
+        vop::CND_GT => cnd4(">"),
         vop::FRC => format!("fract({a})"),
+        vop::TRUNC => format!("trunc({a})"),
         vop::FLOOR => format!("floor({a})"),
+        vop::MAX4 => format!("vec4<f32>(max(max({a}.x,{a}.y), max({a}.z,{a}.w)))"),
+        // dst = (1, src0.y*src1.y, src0.z, src1.w)  (canary kDst)
+        vop::DST => format!("vec4<f32>(1.0, {a}.y * {b}.y, {a}.z, {b}.w)"),
         _ => return None,
     };
     Some(s)
 }
 
 fn scalar_expr(op: u8, a: &str, b: &str, prev: &str) -> Option<String> {
+    // Semantics mirror the runtime interpreter's `exec_scalar_op`
+    // (`shaders/xenos_interp.wgsl`) / canary's `AluScalarOpcode`
+    // (ucode.h:1001+). Side-effecting ops (setp*, kills*, maxas*) need
+    // per-invocation predicate/kill/address state the AOT emitter doesn't
+    // track yet → still `None` (interpreter fallback).
     let s = match op {
         sop::ADDS => format!("({a} + {b})"),
         sop::ADDS_PREV => format!("({a} + {prev})"),
         sop::MULS => format!("({a} * {b})"),
         sop::MULS_PREV => format!("({a} * {prev})"),
+        // muls_prev2 / LIT emulation (canary kMulsPrev2): guard against
+        // -FLT_MAX / non-finite ps & b, and b <= 0.
+        sop::MULS_PREV2 => format!(
+            "select({a} * {prev}, -3.4028235e38, {prev} == -3.4028235e38 || !(\
+             {prev} == {prev}) || abs({prev}) > 3.4028235e38 || !({b} == {b}) || \
+             abs({b}) > 3.4028235e38 || {b} <= 0.0)"
+        ),
         sop::MAXS => format!("max({a}, {b})"),
         sop::MINS => format!("min({a}, {b})"),
-        sop::RCP => format!("xe_rcp({a})"),
+        sop::SEQS => format!("select(0.0, 1.0, {a} == 0.0)"),
+        sop::SGTS => format!("select(0.0, 1.0, {a} > 0.0)"),
+        sop::SGES => format!("select(0.0, 1.0, {a} >= 0.0)"),
+        sop::SNES => format!("select(0.0, 1.0, {a} != 0.0)"),
+        sop::FRCS => format!("fract({a})"),
+        sop::TRUNCS => format!("trunc({a})"),
+        sop::FLOORS => format!("floor({a})"),
+        sop::SUBS => format!("({a} - {b})"),
+        sop::SUBS_PREV => format!("({a} - {prev})"),
+        sop::EXP => format!("exp2({a})"),
+        sop::LOG | sop::LOGC => format!("select(log2({a}), 0.0, {a} == 1.0)"),
+        sop::RCP | sop::RCPC | sop::RCPF => format!("xe_rcp({a})"),
+        sop::RSQ | sop::RSQC | sop::RSQF => {
+            format!("select(0.0, inverseSqrt({a}), {a} > 0.0)")
+        }
+        sop::SQRT => format!("select(0.0, sqrt({a}), {a} >= 0.0)"),
+        sop::SIN => format!("sin({a})"),
+        sop::COS => format!("cos({a})"),
         sop::RETAIN_PREV => prev.to_string(),
         _ => return None,
     };
@@ -528,17 +814,68 @@ mod tests {
     use crate::ucode::alu::{sop, vop};
     use crate::ucode::control_flow::ControlFlowInstruction;
 
+    /// iterate-3T: the real publisher-logo VS (`vs_key 0x03b7b020`, captured
+    /// from the live boot) must now TRANSLATE — pre-3T it rejected with
+    /// `vfetch_fmt` because (a) the `k_16_16` color stream (format 6) was
+    /// unsupported and (b) the export-index model (62=oPos, 0/1=interpolators)
+    /// was a wrong AllocKind heuristic. This locks in the format-6 + per-
+    /// attribute-offset + export-index work so the UV interpolator reaches the
+    /// pixel shader (texcoord in r1) instead of collapsing to a single color.
+    #[test]
+    fn real_logo_vs_translates_with_interpolators() {
+        let ucode: [u32; 30] = [
+            0x70153003, 0x00001200, 0xC2000000, 0x00001006, 0x00001200, 0xC4000000,
+            0x00002007, 0x00002200, 0x00000000, 0x2DF82000, 0x00393A88, 0x00000006,
+            0x05F81000, 0x4006060A, 0x00000306, 0x05F80000, 0x40253FC8, 0x00000406,
+            0xC80F803E, 0x00000000, 0xC2020200, 0xC8038001, 0x00B0B000, 0xC2000000,
+            0xC80F8000, 0x00000000, 0xC2010100, 0x00000000, 0x00000000, 0x00000000,
+        ];
+        let p = crate::ucode::parse_shader(&ucode);
+        let body = match translate(&p, Stage::Vertex) {
+            Translation::Ok(b) => b,
+            Translation::Reject(r) => panic!("logo VS rejected: {r}"),
+        };
+        // Position must come from the export-index-62 path (`opos`) and the
+        // UV/color interpolators must be exported as distinct slots.
+        assert!(body.contains("opos ="), "no position export: {body}");
+        assert!(body.contains("ointerp[0u]"), "no interp0 export: {body}");
+        assert!(body.contains("ointerp[1u]"), "no interp1 export: {body}");
+        // The k_16_16 attribute must unpack via the packed-16 helper.
+        assert!(body.contains("w16"), "no packed-16 unpack for k_16_16: {body}");
+    }
+
+    /// The logo pixel shader (`ps_key 0x03b79001`) samples its texture at the
+    /// interpolated texcoord register r1 — which the PS now seeds from the VS
+    /// interpolator `in.interp1` (Xenos PS-input-GPR mapping). Verifies the UV
+    /// chain so tfetch samples the real UV instead of (0,0).
+    #[test]
+    fn ps_seeds_interpolators_into_registers() {
+        // A trivial PS that just exports — we only assert the preamble wiring.
+        let p = crate::ucode::ParsedShader {
+            cf: vec![ControlFlowInstruction::Exit],
+            instructions: vec![],
+        };
+        let body = match translate(&p, Stage::Pixel) {
+            Translation::Ok(b) => b,
+            Translation::Reject(r) => panic!("trivial PS rejected: {r}"),
+        };
+        assert!(body.contains("r[1] = in.interp1;"), "PS must seed r1 from interp1: {body}");
+    }
+
     fn synthetic_trivial_shader() -> ParsedShader {
         // Single Exec clause: ALU add r0 = r0 + r0; scalar_op = RETAIN_PREV
         // with full write-mask on vector, zero on scalar. Alloc(Position)
         // precedes so the ALU's export (if it were one) would target oPos.
-        // Word-0 bits 29-31 set so all three operands resolve as temps —
-        // matches the prior assertion `r[0u] = (r[0u] + r[0u])`.
-        let w0 = (1u32 << 29) | (1u32 << 30) | (1u32 << 31);
-        let w2 = (vop::ADD as u32)
-            | ((sop::RETAIN_PREV as u32) << 6)
-            | (0xF << 12) // vector_write_mask
-            | (0u32 << 16); // vector_dest = 0
+        // GPUBUG-106 canary layout: dest/mask/scalar_opc in w0; vector_opc +
+        // src_sel in w2. All three operands temps → r0.
+        let w0 = (0u32) // vector_dest = 0
+            | (0xFu32 << 16) // vector_write_mask = 0xF
+            | ((sop::RETAIN_PREV as u32) << 26); // scalar_opc
+        let w1 = 0u32;
+        let w2 = ((vop::ADD as u32) << 24) // vector_opc
+            | (1u32 << 31) // src1_sel = temp
+            | (1u32 << 30) // src2_sel = temp
+            | (1u32 << 29); // src3_sel = temp
         ParsedShader {
             cf: vec![
                 ControlFlowInstruction::Alloc {
@@ -554,7 +891,7 @@ mod tests {
                     predicate_condition: false,
                 },
             ],
-            instructions: vec![w0, 0, w2],
+            instructions: vec![w0, w1, w2],
         }
     }
 
@@ -642,19 +979,17 @@ mod tests {
 
     #[test]
     fn shader_using_c0_emits_xenos_consts_read() {
-        // ALU: r0 = c0 + r0. src_a (low byte) is constant index 0;
-        // src_b (next byte) is temp index 0. src_a_is_temp=false →
-        // src1_sel-style bit at w0 bit 29 = 0; src_b_is_temp=true →
-        // bit 30 = 1. (src_c left as 0/temp; unused.)
-        let w0 = 0x00u32 // src_a = c0
-            | (0x00u32 << 8) // src_b = r0
-            | (0x00u32 << 16) // src_c
-            | (0u32 << 29) // src_a_is_temp = false (constant)
-            | (1u32 << 30); // src_b_is_temp = true (register)
-        let w2 = (vop::ADD as u32)
-            | ((sop::RETAIN_PREV as u32) << 6)
-            | (0xF << 12)
-            | (0u32 << 16);
+        // ALU: r0 = c0 + r0. GPUBUG-106 canary layout. src_a = src1 (w2
+        // 16:23), src_b = src2 (w2 8:15). src1_sel (w2 bit31) = 0 → c0;
+        // src2_sel (w2 bit30) = 1 → r0.
+        let w0 = (0u32) // vector_dest = 0
+            | (0xFu32 << 16) // vector_write_mask
+            | ((sop::RETAIN_PREV as u32) << 26); // scalar_opc
+        let w2 = ((vop::ADD as u32) << 24) // vector_opc
+            | (0u32 << 16) // src1_reg = 0 → c0
+            | (0u32 << 8) // src2_reg = 0 → r0
+            | (0u32 << 31) // src1_sel = 0 (constant)
+            | (1u32 << 30); // src2_sel = 1 (temp)
         let shader = ParsedShader {
             cf: vec![
                 ControlFlowInstruction::Alloc {
@@ -695,9 +1030,16 @@ mod tests {
         let mut ctx = EmitCtx::new(Stage::Vertex);
         let vf = crate::ucode::fetch::VertexFetch {
             fetch_const: 0,
+            const_index_sel: 0,
             src_register: 0,
             dest_register: 0,
             dest_write_mask: 0xF,
+            format: 38, // k_32_32_32_32_FLOAT (4 floats)
+            stride: 4,
+            offset: 0,
+            is_signed: false,
+            is_normalized: true,
+            is_mini_fetch: false,
             raw: [0; 3],
         };
         ctx.emit_vfetch(&vf).expect("emit_vfetch");
@@ -705,6 +1047,70 @@ mod tests {
         assert!(body.contains("gpu_swap("), "emitted vfetch body: {body}");
     }
 
+    fn vf(format: u8, stride: u8, offset: u32, mini: bool) -> crate::ucode::fetch::VertexFetch {
+        crate::ucode::fetch::VertexFetch {
+            fetch_const: 0,
+            const_index_sel: 0,
+            src_register: 0,
+            dest_register: 0,
+            dest_write_mask: 0xF,
+            format,
+            stride,
+            offset,
+            is_signed: false,
+            is_normalized: true,
+            is_mini_fetch: mini,
+            raw: [0; 3],
+        }
+    }
+
+    #[test]
+    fn vfetch_k8888_unpacks_four_channels() {
+        // GPUBUG-112: VertexFormat 6 = k_8_8_8_8 (4× u8 normalized, 1 dword),
+        // NOT k_16_16. All four channels (R,G,B,A) must be unpacked so a
+        // vertex COLOR keeps its blue channel (white texture × white color =
+        // white, not yellow).
+        let mut ctx = EmitCtx::new(Stage::Vertex);
+        ctx.emit_vfetch(&vf(6, 6, 3, false)).expect("emit");
+        let body = ctx.finish();
+        // Four /255.0 channel reads from one packed dword `w16`.
+        assert!(body.contains("let w16 ="), "needs packed dword: {body}");
+        assert_eq!(body.matches("/ 255.0").count(), 4, "four 8-bit channels: {body}");
+    }
+
+    #[test]
+    fn vfetch_mini_inherits_full_stride() {
+        // GPUBUG-114: a vfetch_mini (stride field 0) inherits the stride of the
+        // preceding full vfetch of the same stream (canary ucode.h:733). Emit a
+        // full fetch (stride 7) then a mini fetch and assert the mini indexes by
+        // stride 7, not its tight dword count.
+        let mut ctx = EmitCtx::new(Stage::Vertex);
+        ctx.emit_vfetch(&vf(57, 7, 0, false)).expect("full"); // k_32_32_32_FLOAT
+        ctx.emit_vfetch(&vf(38, 0, 3, true)).expect("mini"); // k_32_32_32_32_FLOAT, mini
+        let body = ctx.finish();
+        assert!(body.contains("vidx * 7u + 3u"), "mini must inherit stride 7: {body}");
+        assert!(!body.contains("vidx * 4u"), "mini must not use tight stride 4: {body}");
+    }
+
+    #[test]
+    fn ps_color_export_is_saturated() {
+        // GPUBUG-115: the PS color export must be clamped to [0,1] (canary
+        // saturates before UNORM RT write) so an out-of-range guest color
+        // doesn't write garbage/white to the sRGB target.
+        let p = crate::ucode::ParsedShader {
+            cf: vec![ControlFlowInstruction::Exit],
+            instructions: vec![],
+        };
+        let body = match translate(&p, Stage::Pixel) {
+            Translation::Ok(b) => b,
+            Translation::Reject(r) => panic!("PS rejected: {r}"),
+        };
+        assert!(
+            body.contains("clamp(ocolor0, vec4<f32>(0.0), vec4<f32>(1.0))"),
+            "PS must saturate color export: {body}"
+        );
+    }
+
     #[test]
     fn loop_clause_rejected() {
         let shader = ParsedShader {
@@ -722,9 +1128,10 @@ mod tests {
 
     #[test]
     fn unsupported_op_rejected() {
-        let w2 = (29u32) // VOP_MAX_A, not in v1 subset
-            | ((sop::RETAIN_PREV as u32) << 6)
-            | (0xF << 12);
+        // GPUBUG-106 layout: vector_write_mask in w0 (16:19), vector_opc in
+        // w2 (24:28). MAX_A (29) is outside the supported subset → reject.
+        let w0 = (0xFu32 << 16) | ((sop::RETAIN_PREV as u32) << 26);
+        let w2 = (29u32) << 24; // VOP_MAX_A
         let shader = ParsedShader {
             cf: vec![ControlFlowInstruction::Exec {
                 address: 0,
@@ -734,7 +1141,7 @@ mod tests {
                 predicated: false,
                 predicate_condition: false,
             }],
-            instructions: vec![0, 0, w2],
+            instructions: vec![w0, 0, w2],
         };
         assert!(matches!(
             translate(&shader, Stage::Vertex),

crates/xenia-gpu/src/ucode/alu.rs

@@ -71,33 +71,50 @@ pub fn decode_alu(words: [u32; 3]) -> AluInstruction {
     let w0 = words[0];
     let w1 = words[1];
     let w2 = words[2];
+    // GPUBUG-106 (iterate-3S): correct the dword field map to match canary's
+    // `AluInstruction` union (ucode.h:2036-2086). Pre-fix this read the
+    // dest/mask/export/scalar-opcode out of `w2`, but they live in `w0`; the
+    // vector opcode + source registers live in `w2`, and swizzle/negate/pred
+    // in `w1`. The misread made every *export* ALU decode with
+    // `vector_write_mask=0` → no oPos/oColor export emitted → the translated VS
+    // collapsed every vertex to the clip origin (degenerate, nothing drawn).
+    //
+    //   w0: vector_dest(0:5) vector_dest_rel(6) abs_constants(7)
+    //       scalar_dest(8:13) scalar_dest_rel(14) export_data(15)
+    //       vector_write_mask(16:19) scalar_write_mask(20:23)
+    //       vector_clamp(24) scalar_clamp(25) scalar_opc(26:31)
+    //   w1: src3_swiz(0:7) src2_swiz(8:15) src1_swiz(16:23)
+    //       src3/2/1_reg_negate(24/25/26) pred_condition(27) is_predicated(28)
+    //   w2: src3_reg(0:7) src2_reg(8:15) src1_reg(16:23)
+    //       vector_opc(24:28) src3/2/1_sel(29/30/31)
+    //
+    // Our (a,b,c) operands map to canary's (src1,src2,src3).
     AluInstruction {
-        vector_opcode: (w2 & 0x3F) as u8,
-        scalar_opcode: ((w2 >> 6) & 0x3F) as u8,
-        vector_dest: ((w2 >> 16) & 0x7F) as u8,
-        scalar_dest: ((w2 >> 24) & 0x7F) as u8,
-        vector_write_mask: ((w2 >> 12) & 0xF) as u8,
-        scalar_write_mask: ((w2 >> 8) & 0xF) as u8,
-        vector_dest_is_export: ((w2 >> 23) & 1) != 0,
-        scalar_src_is_ps: ((w0 >> 26) & 1) != 0,
-        src_a: (w0 & 0xFF) as u8,
-        src_b: ((w0 >> 8) & 0xFF) as u8,
-        src_c: ((w0 >> 16) & 0xFF) as u8,
-        // Word-0 bits 29-31 are the per-operand temp-vs-constant
-        // selector (canary `src3_sel`/`src2_sel`/`src1_sel`,
-        // ucode.h:2078-2086). Our `src_a` is canary's third operand
-        // (low byte of w0), so its selector is bit 29.
-        src_a_is_temp: ((w0 >> 29) & 1) != 0,
-        src_b_is_temp: ((w0 >> 30) & 1) != 0,
-        src_c_is_temp: ((w0 >> 31) & 1) != 0,
-        src_a_swiz: (w1 & 0xFF) as u8,
+        vector_opcode: ((w2 >> 24) & 0x1F) as u8,
+        scalar_opcode: ((w0 >> 26) & 0x3F) as u8,
+        vector_dest: (w0 & 0x3F) as u8,
+        scalar_dest: ((w0 >> 8) & 0x3F) as u8,
+        vector_write_mask: ((w0 >> 16) & 0xF) as u8,
+        scalar_write_mask: ((w0 >> 20) & 0xF) as u8,
+        vector_dest_is_export: ((w0 >> 15) & 1) != 0,
+        // Not a real microcode bit — the scalar pipe selects `ps` implicitly
+        // via the *_PREV opcodes, which `scalar_expr` handles by opcode.
+        scalar_src_is_ps: false,
+        src_a: ((w2 >> 16) & 0xFF) as u8,
+        src_b: ((w2 >> 8) & 0xFF) as u8,
+        src_c: (w2 & 0xFF) as u8,
+        // sel==1 → operand is a temp register; sel==0 → ALU constant.
+        src_a_is_temp: ((w2 >> 31) & 1) != 0,
+        src_b_is_temp: ((w2 >> 30) & 1) != 0,
+        src_c_is_temp: ((w2 >> 29) & 1) != 0,
+        src_a_swiz: ((w1 >> 16) & 0xFF) as u8,
         src_b_swiz: ((w1 >> 8) & 0xFF) as u8,
-        src_c_swiz: ((w1 >> 16) & 0xFF) as u8,
-        src_a_negate: ((w1 >> 24) & 1) != 0,
+        src_c_swiz: (w1 & 0xFF) as u8,
+        src_a_negate: ((w1 >> 26) & 1) != 0,
         src_b_negate: ((w1 >> 25) & 1) != 0,
-        src_c_negate: ((w1 >> 26) & 1) != 0,
-        predicated: ((w0 >> 27) & 1) != 0,
-        predicate_condition: ((w0 >> 28) & 1) != 0,
+        src_c_negate: ((w1 >> 24) & 1) != 0,
+        predicated: ((w1 >> 28) & 1) != 0,
+        predicate_condition: ((w1 >> 27) & 1) != 0,
         raw: words,
     }
 }
@@ -225,19 +242,24 @@ mod tests {
 
     #[test]
     fn decode_extracts_opcodes_and_dests() {
-        // Build a minimal ALU word:
-        //   vector_opcode = ADD (0), scalar_opcode = RCP (22),
-        //   vector_dest = 3, scalar_dest = 7, vector_write_mask = 0xF
-        let w2 = (vop::ADD as u32)
-            | ((sop::RCP as u32) << 6)
-            | (0xF << 12) // vector_write_mask
-            | (3u32 << 16) // vector_dest
-            | (7u32 << 24); // scalar_dest
-        let alu = decode_alu([0, 0, w2]);
+        // GPUBUG-106: correct canary field map. w0 carries dest/mask/scalar_opc;
+        // w2 carries vector_opc + source regs.
+        //   vector_opcode = ADD (0) → w2 bits 24:28
+        //   scalar_opcode = RCP (22) → w0 bits 26:31
+        //   vector_dest = 3 → w0 bits 0:5, scalar_dest = 7 → w0 bits 8:13
+        //   vector_write_mask = 0xF → w0 bits 16:19, export_data → w0 bit 15
+        let w0 = 3u32 // vector_dest
+            | (7u32 << 8) // scalar_dest
+            | (1u32 << 15) // export_data
+            | (0xFu32 << 16) // vector_write_mask
+            | ((sop::RCP as u32) << 26); // scalar_opc
+        let w2 = (vop::ADD as u32) << 24; // vector_opc
+        let alu = decode_alu([w0, 0, w2]);
         assert_eq!(alu.vector_opcode, vop::ADD);
         assert_eq!(alu.scalar_opcode, sop::RCP);
         assert_eq!(alu.vector_dest, 3);
         assert_eq!(alu.scalar_dest, 7);
         assert_eq!(alu.vector_write_mask, 0xF);
+        assert!(alu.vector_dest_is_export);
     }
 }

crates/xenia-gpu/src/ucode/control_flow.rs

@@ -43,7 +43,15 @@ pub enum ControlFlowInstruction {
     Return,
     /// `kAlloc` — pre-allocate export registers (position, interpolators, colors).
     Alloc { size: u32, kind: AllocKind },
-    /// Exit the shader (terminal).
+    /// `kNop` — fills space in the CF block; executes nothing, does not end
+    /// the shader. (Xenos opcode 0.)
+    Nop,
+    /// `kMarkVsFetchDone` — hint that no more vertex fetches will be performed.
+    /// (Xenos opcode 15.) Non-terminating.
+    MarkVsFetchDone,
+    /// Exit the shader (terminal). Synthesized — Xenos has no dedicated exit
+    /// opcode; the shader ends after an `Exec`/`CondExec` clause with the
+    /// END bit set (`is_end`). Retained for callers/tests that reference it.
     Exit,
     /// Unknown / unhandled opcode.
     Unknown { opcode: u8 },
@@ -88,42 +96,66 @@ pub fn decode_cf_pair(word0: u32, word1: u32, word2: u32) -> (ControlFlowInstruc
 fn decode_single(payload: u64) -> ControlFlowInstruction {
     // Top 4 bits of the 48-bit payload.
     let opcode = ((payload >> 44) & 0xF) as u8;
-    // Predicate bit + condition live at the 28..30 range for exec/jmp. Rough
-    // extraction — good enough for the interpreter, which logs unknowns.
-    let predicated = ((payload >> 28) & 1) != 0;
-    let predicate_condition = ((payload >> 29) & 1) != 0;
 
+    // GPUBUG-103 (iterate-3P): clause-level predication is determined by the
+    // *opcode*, not by free bits. The 48-bit CF payload is word0 = bits 0..31,
+    // word1 = bits 32..47. Per canary `ucode.h`:
+    //   * `ControlFlowExecInstruction` (kExec/kExecEnd, opcodes 1/2): NOT
+    //     predicate-gated — it runs unconditionally.
+    //   * `ControlFlowCondExecInstruction` (kCondExec/kCondExecEnd, 3/4): gated
+    //     by a *bool constant*, `condition_` at word1 bit 10 = payload bit 42.
+    //     We don't model bool-constant gating in the WGSL paths (the bool is
+    //     virtually always set for these), so treat as unconditional.
+    //   * `ControlFlowCondExecPredInstruction` (kCondExecPred/...End/Clean...,
+    //     5/6/13/14): gated by the *predicate register*; `condition_` at word1
+    //     bit 9 = payload bit 41.
+    // The prior code read bits 28/29 (which fall inside `sequence_`/`vc_hi_`)
+    // and stamped `predicated=true` on plenty of plain `kExec` clauses — which
+    // made the P7 translator reject EVERY splash VS as `cf_cond`, forcing the
+    // interpreter (placeholder geometry) for all draws.
+    let is_pred_gated = matches!(opcode, 5 | 6 | 13 | 14);
+    let predicated = is_pred_gated;
+    let predicate_condition = is_pred_gated && ((payload >> 41) & 1) != 0;
+
+    // Xenos `ControlFlowOpcode` (canary `ucode.h:86-160`):
+    //   0 kNop, 1 kExec, 2 kExecEnd, 3 kCondExec, 4 kCondExecEnd,
+    //   5 kCondExecPred, 6 kCondExecPredEnd, 7 kLoopStart, 8 kLoopEnd,
+    //   9 kCondCall, 10 kReturn, 11 kCondJmp, 12 kAlloc,
+    //   13 kCondExecPredClean, 14 kCondExecPredCleanEnd, 15 kMarkVsFetchDone.
+    // All exec variants share the address(12)/count(3)/sequence(12) layout
+    // of `ControlFlowExecInstruction`; the `*End` variants terminate the
+    // shader. (Prior table was off-by-one — it mapped 0→Exec and 1→Exit,
+    // so a real `kExec` clause was misread as a terminal `Exit`, truncating
+    // the CF block and dropping every `tfetch` in it.)
+    let exec = |is_end: bool| ControlFlowInstruction::Exec {
+        address: (payload & 0xFFF) as u32,
+        count: ((payload >> 12) & 0x7) as u32,
+        sequence: ((payload >> 16) & 0xFFF) as u32,
+        is_end,
+        predicated,
+        predicate_condition,
+    };
     match opcode {
-        0 => ControlFlowInstruction::Exec {
-            address: (payload & 0xFFF) as u32,
-            count: ((payload >> 12) & 0x7) as u32,
-            sequence: ((payload >> 16) & 0xFFF) as u32,
-            is_end: false,
-            predicated,
-            predicate_condition,
-        },
-        1 => ControlFlowInstruction::Exit,
-        2 => ControlFlowInstruction::Exec {
-            address: (payload & 0xFFF) as u32,
-            count: ((payload >> 12) & 0x7) as u32,
-            sequence: ((payload >> 16) & 0xFFF) as u32,
-            is_end: true,
-            predicated,
-            predicate_condition,
-        },
-        6 => ControlFlowInstruction::LoopStart {
+        0 => ControlFlowInstruction::Nop,
+        1 => exec(false),
+        2 => exec(true),
+        3 => exec(false),
+        4 => exec(true),
+        5 => exec(false),
+        6 => exec(true),
+        7 => ControlFlowInstruction::LoopStart {
             address: (payload & 0x3FF) as u32,
             loop_id: ((payload >> 16) & 0x1F) as u32,
         },
-        7 => ControlFlowInstruction::LoopEnd {
+        8 => ControlFlowInstruction::LoopEnd {
             address: (payload & 0x3FF) as u32,
             loop_id: ((payload >> 16) & 0x1F) as u32,
         },
-        8 => ControlFlowInstruction::CondCall {
+        9 => ControlFlowInstruction::CondCall {
             target: (payload & 0x3FF) as u32,
         },
-        9 => ControlFlowInstruction::Return,
-        10 => ControlFlowInstruction::CondJmp {
+        10 => ControlFlowInstruction::Return,
+        11 => ControlFlowInstruction::CondJmp {
             target: (payload & 0x3FF) as u32,
             predicated,
             predicate_condition,
@@ -132,6 +164,9 @@ fn decode_single(payload: u64) -> ControlFlowInstruction {
             size: (payload & 0x7) as u32,
             kind: AllocKind::from_bits(((payload >> 4) & 0x7) as u32),
         },
+        13 => exec(false),
+        14 => exec(true),
+        15 => ControlFlowInstruction::MarkVsFetchDone,
         other => ControlFlowInstruction::Unknown { opcode: other },
     }
 }
@@ -141,12 +176,49 @@ mod tests {
     use super::*;
 
     #[test]
-    fn opcode_exit_decodes() {
-        // opcode 1 (Exit) in bits 44..47 of A's 48-bit payload.
+    fn opcode_nop_and_exec_decode() {
+        // Xenos opcode 0 = kNop (non-terminating padding).
+        let payload: u64 = 0u64 << 44;
+        let (hi, lo) = ((payload & 0xFFFF_FFFF) as u32, ((payload >> 32) & 0xFFFF) as u32);
+        assert_eq!(decode_cf_pair(hi, lo, 0).0, ControlFlowInstruction::Nop);
+        // Xenos opcode 1 = kExec (executes instructions; NOT a terminal exit).
         let payload: u64 = 1u64 << 44;
         let (hi, lo) = ((payload & 0xFFFF_FFFF) as u32, ((payload >> 32) & 0xFFFF) as u32);
-        let cf = decode_cf_pair(hi, lo, 0).0;
-        assert_eq!(cf, ControlFlowInstruction::Exit);
+        match decode_cf_pair(hi, lo, 0).0 {
+            ControlFlowInstruction::Exec { is_end, .. } => assert!(!is_end),
+            other => panic!("opcode 1 should be non-end Exec, got {other:?}"),
+        }
+        // Xenos opcode 15 = kMarkVsFetchDone (non-terminating hint).
+        let payload: u64 = 15u64 << 44;
+        let (hi, lo) = ((payload & 0xFFFF_FFFF) as u32, ((payload >> 32) & 0xFFFF) as u32);
+        assert_eq!(
+            decode_cf_pair(hi, lo, 0).0,
+            ControlFlowInstruction::MarkVsFetchDone
+        );
+    }
+
+    #[test]
+    fn real_logo_shader_has_tfetch_clauses() {
+        // The publisher-logo pixel shader E59B2B3DA4AA9008 (captured from the
+        // canary oracle, byte-identical to the microcode our guest IM_LOADs).
+        // Regression for iterate-3M: the old off-by-one opcode table decoded
+        // its leading `kExec` (opcode 1) as a terminal `Exit`, truncating the
+        // CF block so the `tfetch2D` never appeared → flat splash.
+        let ucode: [u32; 24] = [
+            0x00011002, 0x00001200, 0xC4000000, 0x00004003, 0x00002200, 0x00000000,
+            0x10082021, 0x1F1FF688, 0x00004000, 0xC8080001, 0x001B1B00, 0xC1020000,
+            0xC8070000, 0x00C0C000, 0xC1020000, 0xC8070001, 0x00C01B00, 0xC1000100,
+            0xC80F8000, 0x00000000, 0xC2010100, 0x00000000, 0x00000000, 0x00000000,
+        ];
+        let p = crate::ucode::parse_shader(&ucode);
+        let exec_clauses = p
+            .cf
+            .iter()
+            .filter(|c| matches!(c, ControlFlowInstruction::Exec { .. }))
+            .count();
+        assert!(exec_clauses >= 1, "expected >=1 Exec clause, cf={:?}", p.cf);
+        let slots = crate::shader_metrics::tfetch_slots(&p);
+        assert!(!slots.is_empty(), "expected tfetch slots, got none; cf={:?}", p.cf);
     }
 
     #[test]

crates/xenia-gpu/src/ucode/fetch.rs

@@ -17,17 +17,64 @@ pub enum FetchInstruction {
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub struct VertexFetch {
-    /// Vertex fetch constant index (0..=95).
+    /// Vertex fetch *const_index* (5 bits, w0[20:24]). The full fetch-constant
+    /// index is `const_index * 3 + const_index_sel` (canary `ucode.h:700`); use
+    /// [`VertexFetch::const_reg_offset`] for the register-region dword offset.
     pub fetch_const: u8,
+    /// iterate-3X (GPUBUG-110): `const_index_sel` (2 bits, w0[25:26]) — selects
+    /// one of the 3 two-dword vertex-fetch constants packed in each 6-dword
+    /// register group. Dropping this read sub-slot 0 of the group, missing the
+    /// real vertex-buffer base for shaders that use sub-slot 1/2 (the publisher
+    /// logo uses `const_index=31, sel=2`).
+    pub const_index_sel: u8,
     /// Source register index (vertex index in r#).
     pub src_register: u8,
     /// Destination register for the fetched value.
     pub dest_register: u8,
     /// 4-bit write mask.
     pub dest_write_mask: u8,
+    /// iterate-3S (GPUBUG-107): `xenos::VertexFormat` (6 bits, dword1[16:21]).
+    /// Determines how many components to read and their packing. Pre-fix the
+    /// translator hardcoded `k_32_32_32_32_FLOAT` (4 floats, stride 4),
+    /// over-striding 2-float UI quads (`k_32_32_FLOAT`) → wrong/clipped
+    /// positions (the next vertex's X bled into .w, giving negative W → the
+    /// whole rectangle was clipped behind the camera).
+    pub format: u8,
+    /// Dword stride between consecutive vertices (dword2[0:7]).
+    pub stride: u8,
+    /// iterate-3T: dword offset of THIS attribute within the vertex stride
+    /// (dword2[16:38] in canary's `VertexFetchInstruction`; the low 23 bits).
+    /// A 6-dword vertex with position@0 + UV@2 + extra@3 needs this so the
+    /// three vfetches sharing one fetch-constant read different attributes
+    /// instead of all reading offset 0.
+    pub offset: u32,
+    /// `is_signed` = canary `fomat_comp_all`, word1 bit 12 (ucode.h:757) —
+    /// selects signed vs unsigned interpretation of packed integer formats.
+    /// (GPUBUG-113: was read from word1 bit 24, which is inside `exp_adjust`.)
+    pub is_signed: bool,
+    /// `is_normalized` = canary `num_format_all == 0`, word1 bit 13
+    /// (ucode.h:758). Set bit ⇒ integer (un-normalized); clear ⇒ normalized.
+    /// We store the normalized sense directly. (GPUBUG-113: was word1 bit 25.)
+    pub is_normalized: bool,
+    /// `is_mini_fetch` = canary word1 bit 30 (ucode.h:764). A mini-fetch reuses
+    /// the address AND STRIDE of the preceding full vfetch of the same stream;
+    /// its own `stride` field is 0. Required so a vfetch_mini color attribute
+    /// indexes by the real vertex stride instead of its tight dword count.
+    pub is_mini_fetch: bool,
     pub raw: [u32; 3],
 }
 
+impl VertexFetch {
+    /// Dword offset of this fetch's 2-dword constant within the fetch-constant
+    /// register region (`CONST_BASE_FETCH`). Vertex fetch constants are packed
+    /// 3 per 6-dword group: `const_index * 6 + const_index_sel * 2`
+    /// (canary `ucode.h:700` `fetch_constant_index = const_index*3 + sel`,
+    /// each constant 2 dwords).
+    pub fn const_reg_offset(&self) -> u32 {
+        self.fetch_const as u32 * 6 + self.const_index_sel as u32 * 2
+    }
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub struct TextureFetch {
     /// Texture fetch constant index (0..=31).
@@ -54,23 +101,47 @@ pub mod op {
 }
 
 pub fn decode_fetch(words: [u32; 3]) -> FetchInstruction {
+    // Fetch dword0 bitfields (Xenos `ucode.h:740-749` vfetch / `844-845`
+    // tfetch): opcode_value:5, src_reg:6, src_reg_am:1, dst_reg:6,
+    // dst_reg_am:1, (fetch_valid_only|must_be_one):1, const_index:5 @ bit20,
+    // ... The prior decoder read `const_index` from bit 5 (which is actually
+    // `src_reg`), so every fetch reported the wrong fetch-constant slot — the
+    // logo `tfetch2D ..., tf0` was read as `tf1`, and slot 1's empty constant
+    // failed to decode → no texture. The texture-fetch `dimension` lives in
+    // dword2 bits 14..15, not dword1.
     let w0 = words[0];
     let w1 = words[1];
+    let w2 = words[2];
     let opcode = (w0 & 0x1F) as u8;
     match opcode {
         op::VERTEX_FETCH => FetchInstruction::Vertex(VertexFetch {
-            fetch_const: ((w0 >> 5) & 0x1F) as u8,
-            src_register: ((w0 >> 17) & 0x7F) as u8,
-            dest_register: ((w0 >> 10) & 0x7F) as u8,
-            dest_write_mask: ((w1 >> 23) & 0xF) as u8,
+            fetch_const: ((w0 >> 20) & 0x1F) as u8,
+            const_index_sel: ((w0 >> 25) & 0x3) as u8,
+            src_register: ((w0 >> 5) & 0x3F) as u8,
+            dest_register: ((w0 >> 12) & 0x3F) as u8,
+            dest_write_mask: (w1 & 0xF) as u8,
+            // dword1[16:21] = VertexFormat. dword2: stride[0:7],
+            // offset (in dwords) [8:?] — empirically the attribute offset of
+            // the textured logo VS lands in dword2[8:15] (pos@4, UV@3,
+            // 3-float@0 in a 6-dword vertex). signed/normalized live higher.
+            format: ((w1 >> 16) & 0x3F) as u8,
+            stride: (w2 & 0xFF) as u8,
+            offset: (w2 >> 8) & 0xFF,
+            // GPUBUG-113: canary ucode.h:757-758,764 — signed=fomat_comp_all
+            // (w1 bit12), normalized=(num_format_all==0) (w1 bit13),
+            // mini-fetch=(w1 bit30). The previous bit24/25 reads landed inside
+            // `exp_adjust`, so signedness/normalization were effectively random.
+            is_signed: ((w1 >> 12) & 1) != 0,
+            is_normalized: ((w1 >> 13) & 1) == 0,
+            is_mini_fetch: ((w1 >> 30) & 1) != 0,
             raw: words,
         }),
         op::TEXTURE_FETCH => FetchInstruction::Texture(TextureFetch {
-            fetch_const: ((w0 >> 5) & 0x1F) as u8,
-            src_register: ((w0 >> 17) & 0x7F) as u8,
-            dest_register: ((w0 >> 10) & 0x7F) as u8,
-            dest_write_mask: ((w1 >> 23) & 0xF) as u8,
-            dimension: ((w1 >> 29) & 0x3) as u8,
+            fetch_const: ((w0 >> 20) & 0x1F) as u8,
+            src_register: ((w0 >> 5) & 0x3F) as u8,
+            dest_register: ((w0 >> 12) & 0x3F) as u8,
+            dest_write_mask: (w1 & 0xF) as u8,
+            dimension: ((w2 >> 14) & 0x3) as u8,
             raw: words,
         }),
         _ => FetchInstruction::Unknown { opcode, raw: words },
@@ -83,8 +154,9 @@ mod tests {
 
     #[test]
     fn decode_vertex_fetch() {
-        // opcode=0 (vertex), fetch_const=5, src=2, dest=7.
-        let w0 = 0u32 | (5 << 5) | (7 << 10) | (2 << 17);
+        // opcode=0 (vertex). Xenos dword0: src_reg@bit5, dst_reg@bit12,
+        // const_index@bit20. fetch_const=5, src=2, dest=7.
+        let w0 = 0u32 | (2 << 5) | (7 << 12) | (5 << 20);
         let v = decode_fetch([w0, 0, 0]);
         match v {
             FetchInstruction::Vertex(vf) => {
@@ -96,13 +168,69 @@ mod tests {
         }
     }
 
+    #[test]
+    fn vertex_fetch_const_index_sel_and_reg_offset() {
+        // iterate-3X (GPUBUG-110): the real publisher-logo vfetch (w0 =
+        // 0x2DF82000) encodes const_index=31, const_index_sel=2. Its fetch
+        // constant lives at dword offset `31*6 + 2*2 = 190` (reg 0x48BE), not
+        // `31*6 = 186` (reg 0x48BA, which held the unused 0x1 slot). Dropping
+        // the sel field made the logo geometry resolve as "no vertex buffer".
+        let v = decode_fetch([0x2DF8_2000, 0, 0]);
+        match v {
+            FetchInstruction::Vertex(vf) => {
+                assert_eq!(vf.fetch_const, 31, "const_index");
+                assert_eq!(vf.const_index_sel, 2, "const_index_sel");
+                assert_eq!(vf.const_reg_offset(), 190, "reg offset = 31*6 + 2*2");
+            }
+            other => panic!("expected Vertex, got {other:?}"),
+        }
+        // sel=0 collapses to the legacy `fetch_const*6` offset (back-compat).
+        let v0 = decode_fetch([0u32 | (5 << 20), 0, 0]);
+        if let FetchInstruction::Vertex(vf) = v0 {
+            assert_eq!(vf.const_index_sel, 0);
+            assert_eq!(vf.const_reg_offset(), 30);
+        } else {
+            panic!("expected Vertex");
+        }
+    }
+
+    #[test]
+    fn vertex_fetch_signed_normalized_mini_bits() {
+        // GPUBUG-113: canary ucode.h:757-758,764 — is_signed=fomat_comp_all
+        // (w1 bit12), is_normalized=(num_format_all==0) (w1 bit13),
+        // is_mini_fetch=(w1 bit30). Validate each bit independently.
+        let mk = |w1: u32| match decode_fetch([0, w1, 0]) {
+            FetchInstruction::Vertex(vf) => vf,
+            _ => panic!("vertex"),
+        };
+        // No bits: unsigned, normalized, full fetch.
+        let v = mk(0);
+        assert!(!v.is_signed);
+        assert!(v.is_normalized);
+        assert!(!v.is_mini_fetch);
+        // bit12 → signed.
+        assert!(mk(1 << 12).is_signed);
+        // bit13 (num_format_all=1) → NOT normalized.
+        assert!(!mk(1 << 13).is_normalized);
+        // bit30 → mini fetch.
+        assert!(mk(1 << 30).is_mini_fetch);
+        // The old (wrong) bits 24/25 must NOT affect signed/normalized.
+        assert!(!mk(1 << 24).is_signed);
+        assert!(mk(1 << 25).is_normalized);
+    }
+
     #[test]
     fn decode_texture_fetch() {
-        let w0 = 1u32 | (3 << 5) | (4 << 10) | (1 << 17);
-        let t = decode_fetch([w0, (2u32 << 29), 0]);
+        // opcode=1 (texture). const_index@bit20=3, src@bit5=1, dst@bit12=4.
+        // dimension lives in dword2 bits 14..15.
+        let w0 = 1u32 | (1 << 5) | (4 << 12) | (3 << 20);
+        let w2 = 2u32 << 14;
+        let t = decode_fetch([w0, 0, w2]);
         match t {
             FetchInstruction::Texture(tf) => {
                 assert_eq!(tf.fetch_const, 3);
+                assert_eq!(tf.src_register, 1);
+                assert_eq!(tf.dest_register, 4);
                 assert_eq!(tf.dimension, 2);
             }
             other => panic!("expected Texture, got {other:?}"),

crates/xenia-gpu/src/ucode/mod.rs

@@ -48,6 +48,9 @@ pub mod cf_kind {
     pub const COND_JMP: u32 = 6;
     pub const COND_CALL: u32 = 7;
     pub const RETURN: u32 = 8;
+    /// Non-executing CF clause: `kNop` padding or `kMarkVsFetchDone` hint.
+    /// The WGSL CF walker treats this as a no-op (advance, do not reject).
+    pub const NOP: u32 = 9;
     pub const UNKNOWN: u32 = 15;
 }
 
@@ -136,6 +139,7 @@ fn encode_cf(c: ControlFlowInstruction) -> (u32, u32, u32) {
         }
         CondCall { target } => (cf_kind::COND_CALL, target, 0),
         Return => (cf_kind::RETURN, 0, 0),
+        Nop | MarkVsFetchDone => (cf_kind::NOP, 0, 0),
         Unknown { opcode } => (cf_kind::UNKNOWN, opcode as u32, 0),
     }
 }
@@ -164,9 +168,11 @@ pub struct ParsedShader {
 }
 
 /// Decode a shader blob. `raw_dwords` is a host-endian slice of the entire
-/// microcode buffer (control flow + instructions). Heuristic: CF dword count
-/// is encoded in the first word's low 12 bits of the last exec clause —
-/// canary iterates until it hits a clause of kind `Exit`. We do the same.
+/// microcode buffer (control flow + instructions). The CF block is implicitly
+/// bounded: we walk clause-pair rows until one terminates the shader (an
+/// `Exec`/`CondExec` clause with the END bit set, per Xenos). Everything after
+/// that row is the instruction block; exec/loop addresses are then rebased to
+/// be relative to it.
 pub fn parse_shader(raw_dwords: &[u32]) -> ParsedShader {
     let mut cf = Vec::new();
     // CF clauses are 48-bit (word1 lo 16 + word0 = 48 or so per canary's
@@ -175,22 +181,50 @@ pub fn parse_shader(raw_dwords: &[u32]) -> ParsedShader {
     while i + 2 < raw_dwords.len() {
         let a = decode_cf_pair(raw_dwords[i], raw_dwords[i + 1], raw_dwords[i + 2]);
         let (first, second) = a;
-        let seen_exit = matches!(
-            first,
-            ControlFlowInstruction::Exit | ControlFlowInstruction::Unknown { .. }
-        ) || matches!(
-            second,
-            ControlFlowInstruction::Exit | ControlFlowInstruction::Unknown { .. }
-        );
+        // The CF block ends after the clause that terminates the shader: an
+        // `Exec` with the END bit set (Xenos `kExecEnd`/`kCondExec*End`), a
+        // synthetic `Exit`, or an `Unknown` opcode (decode ran off the CF
+        // block into instruction data — stop defensively). `Nop` padding
+        // does NOT terminate. (Previously this stopped on the first `Exit`,
+        // but with the corrected opcode table opcode 1 is `kExec`, not exit,
+        // so real exec clauses kept the parse going as intended.)
+        let terminates = |cf: &ControlFlowInstruction| {
+            matches!(
+                cf,
+                ControlFlowInstruction::Exec { is_end: true, .. }
+                    | ControlFlowInstruction::Exit
+                    | ControlFlowInstruction::Unknown { .. }
+            )
+        };
+        let seen_end = terminates(&first) || terminates(&second);
         cf.push(first);
         cf.push(second);
         i += 3;
-        if seen_exit {
+        if seen_end {
             break;
         }
     }
     // Everything after `i` dwords is the instruction block.
     let instructions = raw_dwords[i..].to_vec();
+    // Xenos exec/loop `address` fields are absolute instruction-triple indices
+    // counted from shader dword 0, but `instructions` here begins *after* the
+    // CF block. Rebase those addresses to be relative to the instruction block
+    // (subtract the CF triple count) so `address * 3` indexes `instructions`
+    // directly. (Without this, every exec read 3 dwords too far per CF triple —
+    // the publisher-logo `tfetch` triple was skipped → flat splash.)
+    let cf_triples = (i / 3) as u32;
+    for clause in cf.iter_mut() {
+        match clause {
+            ControlFlowInstruction::Exec { address, .. } => {
+                *address = address.saturating_sub(cf_triples);
+            }
+            ControlFlowInstruction::LoopStart { address, .. }
+            | ControlFlowInstruction::LoopEnd { address, .. } => {
+                *address = address.saturating_sub(cf_triples);
+            }
+            _ => {}
+        }
+    }
     ParsedShader { cf, instructions }
 }
 
@@ -235,15 +269,19 @@ mod tests {
     }
 
     #[test]
-    fn trivial_exit_clause_stops_parsing() {
-        // Two clauses: [NOP (kind=0), EXIT (kind=1)] encoded per canary.
-        // Exit clause is opcode 1 in the top 4 bits of the upper 16 bits.
-        let w0 = 0u32; // clause A body
-        let w1 = (1u32 << 12) << 16; // upper 16 bits = 0x1000 → opcode=1 (EXIT) for clause A
-        let w2 = 0u32;
-        let p = parse_shader(&[w0, w1, w2, 0xDEAD_BEEF]);
+    fn exec_end_clause_stops_parsing() {
+        // Row: clause B = kExecEnd (opcode 2) terminates the CF block.
+        // 48-bit payload of B occupies hi16(word1) + word2; opcode lives in
+        // bits 44..47 of that payload. Put opcode 2 there: payload bit 44 set
+        // for the `2` → (2 << 44). In B's framing, bits 16..47 come from
+        // word2, so word2 bit (44-16)=28 region holds the opcode nibble.
+        let b_payload: u64 = 2u64 << 44; // kExecEnd
+        // B = lo16 from hi16(word1), hi from word2. Reconstruct word1/word2.
+        let word1 = ((b_payload & 0xFFFF) as u32) << 16; // B's low 16 bits → hi16(word1)
+        let word2 = ((b_payload >> 16) & 0xFFFF_FFFF) as u32;
+        let p = parse_shader(&[0, word1, word2, 0xDEAD_BEEF]);
         assert!(!p.cf.is_empty());
-        // Exit detected → remaining dword is instruction data.
+        // ExecEnd detected in the first row → remaining dword is instruction data.
         assert_eq!(p.instructions, vec![0xDEAD_BEEF]);
     }
 }

crates/xenia-kernel/src/exports.rs

@@ -1652,6 +1652,79 @@ fn nt_set_information_file(ctx: &mut PpcContext, mem: &GuestMemory, state: &mut
         return;
     }
 
+    // XFileRenameInformation (10): move the backing file to a new path.
+    // Sylpheed's asset-cache decompresses each packed resource to a staging
+    // `cache:\<hash><tail>.tmp` then renames it into its final nested path
+    // `cache:\<hash>\<dir>\<file>`. Without an actual host-FS rename the
+    // nested target stays empty, the later read-back of the decompressed
+    // asset (e.g. the title logo texture `\69d8e45c\e\534ffea`) misses, and
+    // the logo never loads. Mirror canary `xboxkrnl_io_info.cc:226`
+    // (`X_FILE_RENAME_INFORMATION{ replace_existing@0, root_dir_handle@4,
+    // ansi_string@8 }` → `file->Rename(TranslateAnsiPath(ansi_string))`).
+    if info_class == 10 {
+        // Read the target path from the embedded ANSI_STRING at info_ptr+8.
+        let target_raw = match crate::path::read_ansi_string(mem, info_ptr + 8) {
+            Some(s) if !s.is_empty() => s,
+            _ => {
+                const STATUS_OBJECT_NAME_INVALID: u64 = 0xC000_0033;
+                ctx.gpr[3] = STATUS_OBJECT_NAME_INVALID;
+                return;
+            }
+        };
+        // Resolve the destination against the host cache backing dir. We only
+        // support renames within the writable `cache:` mount (the only place
+        // a guest can create files); disc/synth entries are read-only.
+        let new_host = state.resolve_cache_path(&target_raw);
+        // Current backing host path of the handle.
+        let old_host = match state.objects.get(&handle) {
+            Some(KernelObject::File { host_path: Some(hp), .. }) => Some(hp.clone()),
+            Some(KernelObject::File { .. }) => None,
+            _ => {
+                ctx.gpr[3] = STATUS_INVALID_HANDLE;
+                return;
+            }
+        };
+        let status: u64 = match (old_host, new_host) {
+            (Some(old), Some(new)) => {
+                if let Some(parent) = new.parent() {
+                    let _ = std::fs::create_dir_all(parent);
+                }
+                match std::fs::rename(&old, &new) {
+                    Ok(()) => {
+                        // Update the handle so subsequent I/O targets the new
+                        // host path + guest path.
+                        if let Some(KernelObject::File { path, host_path, .. }) =
+                            state.objects.get_mut(&handle)
+                        {
+                            *path = crate::path::normalize_path(&target_raw);
+                            *host_path = Some(new.clone());
+                        }
+                        tracing::info!(
+                            "NtSetInformationFile rename cache {:?} -> {:?} ({:?})",
+                            old, new, target_raw
+                        );
+                        STATUS_SUCCESS
+                    }
+                    Err(e) => {
+                        tracing::warn!(
+                            "NtSetInformationFile rename {:?} -> {:?} failed: {}",
+                            old, new, e
+                        );
+                        STATUS_UNSUCCESSFUL
+                    }
+                }
+            }
+            // Non-cache (read-only VFS) source/target: acknowledge without a
+            // host move, matching the prior permissive behaviour.
+            _ => STATUS_SUCCESS,
+        };
+        if iosb_ptr != 0 {
+            write_io_status_block(mem, iosb_ptr, status as u32, info_length);
+        }
+        ctx.gpr[3] = status;
+        return;
+    }
+
     // Handle lookup.
     let Some(KernelObject::File { size, position, host_path, .. }) = state.objects.get_mut(&handle) else {
         ctx.gpr[3] = STATUS_INVALID_HANDLE;
@@ -2883,10 +2956,12 @@ fn vd_initialize_ring_buffer(ctx: &mut PpcContext, _mem: &GuestMemory, state: &m
     // packets directly into ring memory at the current WPTR (the GPU
     // backend lives on a worker thread under `--gpu-thread` so we can't
     // read its `ring.base` from the kernel side without a channel hop).
-    // Per canary: size_log2 is log2(size in BYTES), so size in dwords =
-    // 2^size_log2 / 4 = 1 << (size_log2 - 2).
+    // Per canary `CommandProcessor::InitializeRingBuffer`: the ring is
+    // `1 << (size_log2 + 3)` bytes = `1 << (size_log2 + 1)` dwords (`r4` is
+    // log2 of the size in quadwords). Kept in sync with
+    // `GpuSystem::initialize_ring_buffer`. (Currently bookkeeping-only.)
     state.ring_base = ptr;
-    state.ring_size_dwords = if size_log2 >= 2 { 1u32 << (size_log2 - 2) } else { 0 };
+    state.ring_size_dwords = 1u32 << (size_log2 + 1);
     ctx.gpr[3] = 0;
 }
 
@@ -2997,53 +3072,87 @@ fn vd_swap(ctx: &mut PpcContext, mem: &GuestMemory, state: &mut KernelState) {
     // xboxkrnl_video.cc:479. Currently skipped (see below).
     let _ = fetch_dwords; // silence unused — will be live again under the deferred path
 
-    // The original M2b path zero-filled buffer_ptr (in the system command
-    // buffer) and bumped WPTR by 64 to expose the game's own ring writes.
-    // Keep that untouched — the game still expects buffer_ptr to be a
-    // skippable scratch area, and the bump still exposes any game-batched
-    // PM4 packets for the drain.
+    // iterate-2V: mirror xenia-canary `VdSwap_entry` (xboxkrnl_video.cc:518-548)
+    // FAITHFULLY. The game reserves 64 dwords (256 bytes) in the primary ring
+    // at `buffer_ptr`; canary writes a `PM4_TYPE0(SHADER_CONSTANT_FETCH_00_0)`
+    // fetch-constant patch followed by `PM4_TYPE3(PM4_XE_SWAP)`, then pads with
+    // NOPs — and **NEVER touches `CP_RB_WPTR`**. The game advances the primary
+    // ring write-pointer itself via its own doorbell once it has finished
+    // populating the reserved slot, so VdSwap only fills the bytes.
+    //
+    // iterate-2V FIX (the bug this removes): a prior revision bumped the
+    // primary ring `CP_RB_WPTR` out-of-band here (`extend_write_ptr_by(64)`).
+    // But `buffer_ptr` (~0x4add6efc) is NOT inside the primary ring (base
+    // ~0x4adcd000, 8192 dwords) — it lives ~10k dwords past it, in the
+    // renderer indirect-buffer region. The bogus WPTR bump pushed the GPU
+    // read-pointer PAST the guest's real write-pointer, the drain treated the
+    // overshoot as a circular wrap, and **re-executed the splash's draw
+    // indirect-buffers ~2×** — inflating draws to 78 (real splash ≈ 28; 12
+    // INDIRECT_BUFFERs vs the real 6). Canary's `VdSwap_entry` writes the
+    // block and returns; the swap-complete CP interrupt comes only from the
+    // game's own in-stream `PM4_INTERRUPT` packets, never from VdSwap.
     if buffer_ptr != 0 {
-        for i in 0..64u32 {
-            mem.write_u32(buffer_ptr + i * 4, xenia_gpu::pm4::make_packet_type2());
+        let mut off = 0u32;
+        let mut put = |i: &mut u32, v: u32| {
+            mem.write_u32(buffer_ptr + *i * 4, v);
+            *i += 1;
+        };
+        // PM4_TYPE0 fetch-constant slot-0 patch (6 dwords payload). The
+        // base_address field is patched to the physical frontbuffer so the
+        // bloom/blur "sample frame N for frame N+1" path reads the right page.
+        let mut patched = fetch_dwords;
+        patched[1] = (patched[1] & 0x0000_0FFF) | ((frontbuffer_addr >> 12) << 12);
+        put(
+            &mut off,
+            xenia_gpu::pm4::make_packet_type0(
+                xenia_gpu::gpu_system::CONST_BASE_FETCH as u16,
+                6,
+            ),
+        );
+        for d in patched {
+            put(&mut off, d);
+        }
+        // PM4_TYPE3(PM4_XE_SWAP, 4 dwords): signature, frontbuffer_phys, w, h.
+        put(
+            &mut off,
+            xenia_gpu::pm4::make_packet_type3(xenia_gpu::pm4::PM4_XE_SWAP, 4),
+        );
+        put(&mut off, xenia_gpu::pm4::SWAP_SIGNATURE);
+        put(&mut off, frontbuffer_addr);
+        put(&mut off, width);
+        put(&mut off, height);
+        // Pad the remainder with NOP (Type-2) packets.
+        while off < 64 {
+            put(&mut off, xenia_gpu::pm4::make_packet_type2());
         }
     }
-    state.gpu.extend_write_ptr_by(64);
+    // NOTE: We deliberately do NOT bump `CP_RB_WPTR` here (see the iterate-2V
+    // comment above). The drain below consumes only the packets the game has
+    // legitimately advanced the write-pointer over.
 
-    // GPUBUG-DRAIN-001: notify the swap directly.
-    //
-    // Per xenia-canary `VdSwap_entry` (xboxkrnl_video.cc:438-521), the
-    // textbook approach is to inject `PM4_TYPE0(SHADER_CONSTANT_FETCH_00_0)`
-    // (fetch-constant slot-0 patch for the Sylpheed bloom/blur "frame N+1"
-    // sample) followed by `PM4_TYPE3(PM4_XE_SWAP)` directly into the
-    // primary ring at WPTR, then let the natural drain consume them.
-    //
-    // That works in **pure lockstep** (drain runs at every kernel callback
-    // boundary, ring has at most a few hundred packets pending). It
-    // **does not** work under `--parallel` (CPU + GPU ring contention) —
-    // observed empirically: vd_swap's `drain_to_current_wptr` consumes
-    // 8-10 million game-batched IB packets in the 900 ms inline-deadline
-    // window without reaching our tail-injected PM4_XE_SWAP. Under
-    // threaded backend the worker has the same deadline. Either:
-    //   (a) the safety-net direct notify (below) fires and gets the swap
-    //       counted — but if the worker *eventually* drains past our
-    //       injected packet later it would double-count,
-    //   (b) we extend the deadline so far that vd_swap blocks for many
-    //       seconds — unreasonable for a kernel callback.
-    //
-    // Skip the ring injection unconditionally and post `notify_xe_swap`
-    // directly. The drain still runs (game packets execute as normal).
-    // **Trade-off**: the slot-0 fetch-constant patch is deferred —
-    // tracked as GPUBUG-FETCH-PATCH-001. Sylpheed currently has draws=0,
-    // so a stale slot 0 has no observable effect.
+    // Drain the ring up to whatever the game has actually submitted; any
+    // in-stream `PM4_INTERRUPT` / draw packets execute in order. The
+    // reserved-slot PM4_XE_SWAP is consumed by the GPU only once the game
+    // advances its own doorbell over it. The swap-counter safety net below
+    // keeps host swap bookkeeping live in the meantime.
     let drained = state.gpu.drain_to_current_wptr(mem);
     tracing::debug!(drained, "VdSwap: drained PM4 packets");
 
-    // Direct swap notification. Inline mode bumps `swaps_seen`
-    // synchronously; threaded mode posts a `GpuCommand::NotifyXeSwap`
-    // and the worker bumps it asynchronously.
+    // Safety net: if the drain did NOT reach our PM4_XE_SWAP this call (e.g.
+    // an undersized inline deadline left game-batched packets pending), still
+    // bump the host swap counter so the UI present + swap stats stay live.
+    // Skip when the in-stream PM4_XE_SWAP already recorded this frontbuffer
+    // (avoids double-counting). This path does NOT raise a CP interrupt.
     if frontbuffer_addr != 0 && width > 0 && height > 0 {
+        let already_swapped = state
+            .gpu
+            .as_inline_mut()
+            .map(|g| g.last_swap.map(|s| s.frontbuffer_phys) == Some(frontbuffer_addr))
+            .unwrap_or(false);
+        if !already_swapped {
             state.gpu.notify_xe_swap(frontbuffer_addr, width, height);
         }
+    }
 
     // The remaining vd_swap work (UI publish: shader blobs, constants,
     // texture cache, frontbuffer detile, ui.notify_swap) reads
@@ -3080,16 +3189,24 @@ fn vd_swap(ctx: &mut PpcContext, mem: &GuestMemory, state: &mut KernelState) {
         );
         ui.publish_assets(blobs, constants);
 
-        // P5: try to decode the primary texture (fetch constant slot 0).
-        // Slot 0 is the convention most games use for their main bound
-        // texture at draw time; full N-slot binding waits for P6+. If the
-        // slot is unset or the format isn't supported (magenta stub kicks
-        // in host-side), we skip.
-        //
-        // Texture fetch constants live at `CONST_BASE_FETCH + slot*6` in
-        // the register file; we read the 6 dwords, decode the key, hit
-        // the CPU cache (with page-version freshness), and clone the
-        // decoded bytes across the bridge.
+        // P5b: publish the texture the last draw's *active pixel shader*
+        // actually sampled. The GPU draw handler decodes the PS's real
+        // `tfetch` fetch-constant slots into `last_draw_textures`; we publish
+        // the first (the UI binds a single texture today). When the last draw
+        // used a flat (no-tfetch) shader the list is empty, so we fall back to
+        // the legacy slot-0 probe to preserve behavior on flat-only frames.
+        // The legacy single-texture `publish_texture` bridge wants
+        // `(TextureKey, bytes)`; `last_draw_textures` now also carries the
+        // content version (for the per-draw host-cache re-upload). Drop it here.
+        let published = gpu_inline
+            .last_draw_textures
+            .first()
+            .map(|(k, _v, b)| (*k, b.clone()))
+            .or_else(|| {
+            // Fallback: probe fetch constant slot 0 directly. Texture fetch
+            // constants live at `CONST_BASE_FETCH + slot*6` in the register
+            // file; read 6 dwords, decode the key, hit the CPU cache with
+            // page-version freshness, clone the bytes across the bridge.
             const TEX_SLOT: u32 = 0;
             let mut fetch6 = [0u32; 6];
             for (i, slot) in fetch6.iter_mut().enumerate() {
@@ -3097,10 +3214,9 @@ fn vd_swap(ctx: &mut PpcContext, mem: &GuestMemory, state: &mut KernelState) {
                     .register_file
                     .read(xenia_gpu::gpu_system::CONST_BASE_FETCH + TEX_SLOT * 6 + i as u32);
             }
-        let published = if let Some(key) = xenia_gpu::texture_cache::decode_fetch_constant(fetch6)
-        {
-            // Span over the entire tiled texture footprint to pick the
-            // max page version covering it.
+            let key = xenia_gpu::texture_cache::decode_fetch_constant(fetch6)?;
+            // Span over the entire tiled texture footprint to pick the max
+            // page version covering it.
             let bi = key.format.block_info();
             let span_bytes = (key.pitch_texels as u32)
                 * (key.height as u32)
@@ -3118,12 +3234,20 @@ fn vd_swap(ctx: &mut PpcContext, mem: &GuestMemory, state: &mut KernelState) {
                     None
                 }
             }
-        } else {
-            None
-        };
+        });
         metrics::gauge!("gpu.texture_cache.entries")
             .set(gpu_inline.texture_cache.len() as f64);
         ui.publish_texture(published);
+
+        // iterate-3O: publish this frame's captured per-draw geometry and
+        // reset the accumulator for the next frame. The UI replays these as
+        // real guest draws (real vertices + prim type) instead of synthetic
+        // placeholder shapes. `frame_captures` is `Some` only under `--ui`.
+        if let Some(caps) = gpu_inline.frame_captures.as_mut() {
+            let drained = std::mem::take(caps);
+            metrics::counter!("gpu.geometry.published").increment(drained.len() as u64);
+            ui.publish_geometry(drained);
+        }
     }
     // Notify the UI.
     if let Some(ui) = state.ui.clone() {
@@ -5547,6 +5671,67 @@ mod tests {
         }
     }
 
+    /// `NtSetInformationFile` class 10 (`XFileRenameInformation`) must move
+    /// the backing host file to the new `cache:` path and update the handle.
+    /// Mirrors Sylpheed's asset-cache `.tmp` → `\<hash>\<dir>\<file>` move;
+    /// without it the nested target stays empty and the decompressed asset
+    /// (logo texture) never reads back. Faithful to canary `file->Rename`.
+    #[test]
+    fn nt_set_information_file_rename_moves_cache_file() {
+        let (mut ctx, mut mem, mut state) = fresh();
+        // Real temp cache root + a staging `.tmp` file with known bytes.
+        let root = std::env::temp_dir().join(format!("xenia-rs-rename-test-{}", std::process::id()));
+        let _ = std::fs::remove_dir_all(&root);
+        std::fs::create_dir_all(&root).unwrap();
+        let old_host = root.join("69d8e45ce534ffea.tmp");
+        std::fs::write(&old_host, b"LOGOTEX!").unwrap();
+        state.cache_root = Some(root.clone());
+        // Open handle whose backing host_path is the staging file.
+        let handle = state.alloc_handle_for(KernelObject::File {
+            path: "69d8e45ce534ffea.tmp".to_string(),
+            size: 8,
+            position: 0,
+            data: Arc::new(Vec::new()),
+            dir_enum_pos: None,
+            host_path: Some(old_host.clone()),
+        });
+        // X_FILE_RENAME_INFORMATION { replace@0, root_dir@4, ANSI_STRING@8 }.
+        // ANSI_STRING { len u16, max u16, buf u32 } at info_ptr+8; buffer holds
+        // the target path "cache:\69d8e45c\e\534ffea".
+        let info_ptr = SCRATCH_BASE + 0x100;
+        let str_buf = SCRATCH_BASE + 0x200;
+        let target = b"cache:\\69d8e45c\\e\\534ffea";
+        for (i, b) in target.iter().enumerate() {
+            mem.write_u8(str_buf + i as u32, *b);
+        }
+        mem.write_u32(info_ptr, 0); // replace_existing
+        mem.write_u32(info_ptr + 4, 0); // root_dir_handle
+        mem.write_u16(info_ptr + 8, target.len() as u16); // ANSI_STRING.Length
+        mem.write_u16(info_ptr + 10, target.len() as u16); // MaximumLength
+        mem.write_u32(info_ptr + 12, str_buf); // Buffer
+        let iosb_ptr = SCRATCH_BASE + 0x140;
+        ctx.gpr[3] = handle as u64;
+        ctx.gpr[4] = iosb_ptr as u64;
+        ctx.gpr[5] = info_ptr as u64;
+        ctx.gpr[6] = 16;
+        ctx.gpr[7] = 10; // XFileRenameInformation
+        nt_set_information_file(&mut ctx, &mut mem, &mut state);
+        assert_eq!(ctx.gpr[3], STATUS_SUCCESS);
+        // Staging file gone; nested target exists with the same bytes.
+        let new_host = root.join("69d8e45c").join("e").join("534ffea");
+        assert!(!old_host.exists(), "staging .tmp should be moved away");
+        assert_eq!(std::fs::read(&new_host).unwrap(), b"LOGOTEX!");
+        // Handle now points at the new host + guest path.
+        match state.objects.get(&handle) {
+            Some(KernelObject::File { host_path: Some(hp), path, .. }) => {
+                assert_eq!(hp, &new_host);
+                assert_eq!(path, "cache:/69d8e45c/e/534ffea");
+            }
+            _ => panic!("file handle lost or host_path missing"),
+        }
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
     /// Read-only VFS — truncating to a different size must fail with
     /// `STATUS_UNSUCCESSFUL`, matching Canary's error path when
     /// `file->SetLength(...)` can't honour the request.

crates/xenia-kernel/src/interrupts.rs

@@ -30,6 +30,12 @@ use xenia_cpu::ThreadRef;
 pub const INTERRUPT_SOURCE_VSYNC: u32 = 0;
 pub const INTERRUPT_SOURCE_CP: u32 = 1;
 
+/// The processor the graphics ISR impersonates for a v-sync interrupt.
+/// Canary hard-codes this: `MarkVblank` → `DispatchInterruptCallback(0, 2)`
+/// (graphics_system.cc:478). CP interrupts instead use the bit index of the
+/// `PM4_INTERRUPT` `cpu_mask`.
+pub const VSYNC_TARGET_CPU: u8 = 2;
+
 /// Guest-registered V-sync / graphics-interrupt callback (from
 /// `VdSetGraphicsInterruptCallback`).
 #[derive(Debug, Clone, Copy)]
@@ -145,9 +151,16 @@ pub type PendingLocalIrq = [std::sync::atomic::AtomicU8;
 pub struct InterruptState {
     /// Registered callback (set by `VdSetGraphicsInterruptCallback`).
     pub callback: Option<GraphicsInterruptCallback>,
-    /// Bounded FIFO of pending interrupt sources awaiting injection.
-    /// Push-back on queue, pop-front on inject. Over-cap pushes drop.
-    pub pending: VecDeque<u32>,
+    /// Bounded FIFO of pending interrupts awaiting injection, as
+    /// `(source, target_cpu)`. Push-back on queue, pop-front on inject.
+    /// Over-cap pushes drop. `target_cpu` is the processor the graphics
+    /// ISR must impersonate (canary `XThread::SetActiveCpu` / the
+    /// `DispatchInterruptCallback(source, cpu)` argument): the bit index
+    /// of the CP `PM4_INTERRUPT` `cpu_mask` for source=1, and a fixed `2`
+    /// for vsync (canary `DispatchInterruptCallback(0, 2)`). The ISR reads
+    /// it from the PCR (`[r13+268]`) to clear the matching per-CPU bit of
+    /// the swap-acknowledge fence.
+    pub pending: VecDeque<(u32, u8)>,
     /// When `Some`, some HW thread is currently running a callback; on
     /// return-to-sentinel we restore this and clear the flag.
     pub saved: Option<SavedCallbackCtx>,
@@ -170,6 +183,28 @@ pub struct InterruptState {
     /// ticker. `tick_vsync_instr` diffs against this to advance
     /// `vsync_accumulator`.
     pub last_instr_count: u64,
+    /// **iterate-3AJ — present-anchored vsync.** Set `true` once the guest
+    /// has presented at least one frame (a `VdSwap`). Before this, the
+    /// vsync ticker uses the legacy fixed instruction-quantum cadence so
+    /// the boot present-loop bootstrap (iterate-2W) still gets the vsyncs
+    /// it needs *before* the first present. After this, vsync is anchored
+    /// to the guest's real present rate (≈1 vblank per present, as on real
+    /// hardware where the title double-buffers at vblank), with only a
+    /// small capped instruction-quantum *fallback* for frames where the
+    /// guest genuinely stops presenting (heavy asset load). This stops the
+    /// proxy from firing ~66 vsyncs during one heavy load frame, which
+    /// collapsed the splash-logo intro fade-in (the guest's vsync counter
+    /// jumped 0→66 in one frame instead of ramping smoothly).
+    pub vsync_present_anchored: bool,
+    /// Last observed guest present (`VdSwap`) count. `tick_vsync_instr`
+    /// diffs the live count against this each call to emit one vblank per
+    /// new present once `vsync_present_anchored` is set.
+    pub last_present_count: u64,
+    /// How many *fallback* (non-present-driven) vsyncs have fired in the
+    /// current dry (no-present) window. Reset to 0 whenever a present
+    /// occurs. Capped at [`DRY_FALLBACK_CAP`] so one heavy non-presenting
+    /// frame cannot fire a long burst of vsyncs (the fade-in regression).
+    pub dry_fallback_fired: u32,
     /// Wall-clock anchor for the production v-sync ticker. `None` until
     /// the first `tick_vsync_wallclock` call (lazy init so unit tests
     /// that never invoke that function don't construct an Instant).
@@ -195,6 +230,21 @@ pub struct InterruptState {
 /// determinism.
 pub const VSYNC_INSTR_PERIOD: u64 = 150_000;
 
+/// **iterate-3AJ — present-anchored vsync fallback.**
+///
+/// Once the guest is in its present loop (`vsync_present_anchored`), each
+/// guest present emits exactly one vblank — vsync *is* the present cadence,
+/// as on real Xbox 360 hardware where the title double-buffers at vblank.
+/// For a frame where the guest stops presenting (e.g. the ~1.1 s splash
+/// asset-load), we still need *some* vsyncs to keep timers / the present
+/// loop alive, but firing one per [`VSYNC_INSTR_PERIOD`] would reproduce the
+/// ~66-vsync spike that collapsed the fade-in. So the fallback fires one
+/// vblank per `VSYNC_INSTR_PERIOD` of *non-presenting* instructions, but at
+/// most [`DRY_FALLBACK_CAP`] per dry window (the counter resets on each
+/// present). A heavy load frame therefore advances the guest vsync counter
+/// by ≤ `DRY_FALLBACK_CAP` (a small ramp like canary's 0/5/10/2/1…), not 66.
+pub const DRY_FALLBACK_CAP: u32 = 4;
+
 /// Wall-clock period for the **production** v-sync ticker. 16.667 ms
 /// targets exactly 60 Hz. KRNBUG-D08 — converting from the
 /// instruction-count proxy fixes the `--parallel` rate drop while
@@ -211,8 +261,9 @@ impl InterruptState {
         });
     }
 
-    /// Queue an interrupt for the next safe injection point.
-    pub fn queue_interrupt(&mut self, source: u32) {
+    /// Queue an interrupt for the next safe injection point. `cpu` is the
+    /// processor the ISR must impersonate (see `pending`).
+    pub fn queue_interrupt(&mut self, source: u32, cpu: u8) {
         if self.callback.is_none() {
             self.dropped += 1;
             return;
@@ -221,37 +272,102 @@ impl InterruptState {
             self.dropped += 1;
             return;
         }
-        self.pending.push_back(source);
+        self.pending.push_back((source, cpu));
     }
 
     /// Peek at the next pending source without removing it.
     pub fn peek_next(&self) -> Option<u32> {
-        self.pending.front().copied()
+        self.pending.front().map(|&(source, _)| source)
+    }
+
+    /// Peek at the target CPU of the next pending interrupt.
+    pub fn peek_next_cpu(&self) -> Option<u8> {
+        self.pending.front().map(|&(_, cpu)| cpu)
     }
 
     /// Pop the next pending source (called by the injector after it has
     /// committed to dispatching it).
     pub fn take_next(&mut self) -> Option<u32> {
-        self.pending.pop_front()
+        self.pending.pop_front().map(|(source, _)| source)
     }
 
-    /// **Legacy** — instruction-count v-sync ticker. Kept for unit tests
-    /// that need a deterministic clock source. Production code calls
-    /// `tick_vsync_wallclock` instead. Returns `true` if at least one
-    /// v-sync was queued.
-    pub fn tick_vsync_instr(&mut self, current_instr_count: u64) -> bool {
+    /// **Present-anchored** instruction-paced v-sync ticker (the lockstep
+    /// production path; also used by unit tests for a deterministic clock).
+    ///
+    /// `current_instr_count` is the running retired-instruction count.
+    /// `present_count` is the guest's running `VdSwap` count (monotonic).
+    ///
+    /// Two regimes:
+    ///
+    /// 1. **Bootstrap** (`!vsync_present_anchored`, i.e. before the guest's
+    ///    first present): legacy fixed-quantum cadence — one vsync per
+    ///    [`VSYNC_INSTR_PERIOD`] retired instructions. The boot present loop
+    ///    (iterate-2W) needs vsyncs delivered *before* it can present, so
+    ///    this regime is unchanged from the original ticker. The first
+    ///    observed present flips `vsync_present_anchored`.
+    ///
+    /// 2. **Present-anchored** (after the first present): one vblank per
+    ///    guest present (vsync *is* the present cadence on real hardware),
+    ///    plus a small capped instruction-quantum fallback ([`DRY_FALLBACK_CAP`]
+    ///    per dry window) so a frame where the guest stops presenting (heavy
+    ///    asset load) still ticks a *few* vsyncs — not ~66, which collapsed
+    ///    the splash fade-in.
+    ///
+    /// Returns `true` if at least one v-sync was queued.
+    pub fn tick_vsync_instr(&mut self, current_instr_count: u64, present_count: u64) -> bool {
         let delta = current_instr_count.saturating_sub(self.last_instr_count);
         self.last_instr_count = current_instr_count;
         self.vsync_accumulator = self.vsync_accumulator.saturating_add(delta);
+
+        let new_presents = present_count.saturating_sub(self.last_present_count);
+        self.last_present_count = present_count;
+        if new_presents > 0 {
+            self.vsync_present_anchored = true;
+        }
+
+        // Regime 1 — bootstrap: legacy fixed instruction quantum. Preserves
+        // the iterate-2W present-loop bootstrap exactly (vsyncs must fire
+        // before the guest can present).
+        if !self.vsync_present_anchored {
             if self.vsync_accumulator < VSYNC_INSTR_PERIOD {
                 return false;
             }
             let periods = self.vsync_accumulator / VSYNC_INSTR_PERIOD;
             self.vsync_accumulator %= VSYNC_INSTR_PERIOD;
             for _ in 0..periods {
-            self.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+                self.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
             }
-        true
+            return true;
+        }
+
+        // Regime 2 — present-anchored.
+        let mut queued = false;
+
+        if new_presents > 0 {
+            // One vblank per guest present. `queue_interrupt` caps the FIFO,
+            // so a burst of presents in one round can't flood. A fresh
+            // present resets the dry-window state.
+            for _ in 0..new_presents {
+                self.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
+            }
+            self.vsync_accumulator = 0;
+            self.dry_fallback_fired = 0;
+            queued = true;
+        } else if self.vsync_accumulator >= VSYNC_INSTR_PERIOD
+            && self.dry_fallback_fired < DRY_FALLBACK_CAP
+        {
+            // Dry frame (no present this tick): the guest stopped presenting
+            // (heavy load). Tick a *capped* number of fallback vsyncs so
+            // timers/the present loop stay alive without re-introducing the
+            // ~66-vsync spike. Consume one period per fired vsync so the
+            // accumulator paces the few fallbacks.
+            self.vsync_accumulator -= VSYNC_INSTR_PERIOD;
+            self.dry_fallback_fired += 1;
+            self.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
+            queued = true;
+        }
+
+        queued
     }
 
     /// **Production** — wall-clock v-sync ticker. Fires
@@ -288,7 +404,7 @@ impl InterruptState {
         self.last_vsync_instant = Some(anchor + advance);
         let to_queue = (periods as usize).min(INTERRUPT_QUEUE_CAP);
         for _ in 0..to_queue {
-            self.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+            self.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
         }
         true
     }
@@ -306,7 +422,7 @@ mod tests {
     #[test]
     fn queue_interrupt_drops_without_callback() {
         let mut s = InterruptState::default();
-        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
         assert_eq!(s.dropped, 1);
         assert!(s.pending.is_empty());
     }
@@ -315,9 +431,9 @@ mod tests {
     fn queue_interrupt_fifo_preserves_order() {
         let mut s = InterruptState::default();
         s.set_callback(0x1000, 0xAB);
-        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
-        s.queue_interrupt(INTERRUPT_SOURCE_CP);
-        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
+        s.queue_interrupt(INTERRUPT_SOURCE_CP, 2);
+        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
         assert_eq!(s.dropped, 0);
         // FIFO: take_next hands them out in push order.
         assert_eq!(s.take_next(), Some(INTERRUPT_SOURCE_VSYNC));
@@ -331,11 +447,11 @@ mod tests {
         let mut s = InterruptState::default();
         s.set_callback(0x1000, 0xAB);
         for _ in 0..INTERRUPT_QUEUE_CAP {
-            s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+            s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
         }
         // Over-cap: drops rather than evicting the oldest.
-        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
-        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC);
+        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
+        s.queue_interrupt(INTERRUPT_SOURCE_VSYNC, VSYNC_TARGET_CPU);
         assert_eq!(s.dropped, 2);
         assert_eq!(s.pending.len(), INTERRUPT_QUEUE_CAP);
     }
@@ -345,9 +461,10 @@ mod tests {
         let mut s = InterruptState::default();
         s.set_callback(0x1000, 0xAB);
         assert_eq!(VSYNC_INSTR_PERIOD, 150_000);
-        assert!(!s.tick_vsync_instr(VSYNC_INSTR_PERIOD - 1));
+        // present_count = 0 → bootstrap regime (legacy fixed quantum).
+        assert!(!s.tick_vsync_instr(VSYNC_INSTR_PERIOD - 1, 0));
         assert!(s.pending.is_empty());
-        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD));
+        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD, 0));
         assert_eq!(s.peek_next(), Some(INTERRUPT_SOURCE_VSYNC));
     }
 
@@ -357,10 +474,59 @@ mod tests {
         // be delivered, not lost.
         let mut s = InterruptState::default();
         s.set_callback(0x1000, 0xAB);
-        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD * 3 + 10));
+        // present_count = 0 → bootstrap regime drains all 3 periods at once.
+        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD * 3 + 10, 0));
         assert_eq!(s.pending.len(), 3);
     }
 
+    #[test]
+    fn tick_vsync_instr_present_anchors_after_first_present() {
+        // iterate-3AJ: once the guest presents, vsync tracks presents (one
+        // vblank per present), NOT the fixed instruction quantum.
+        let mut s = InterruptState::default();
+        s.set_callback(0x1000, 0xAB);
+        // Bootstrap: instruction quantum fires (present_count still 0).
+        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD, 0));
+        assert_eq!(s.pending.len(), 1);
+        let _ = s.take_next();
+        // First present flips to anchored: exactly one vblank for the present.
+        assert!(s.tick_vsync_instr(VSYNC_INSTR_PERIOD * 2, 1));
+        assert!(s.vsync_present_anchored);
+        assert_eq!(s.pending.len(), 1);
+        let _ = s.take_next();
+    }
+
+    #[test]
+    fn tick_vsync_instr_heavy_dry_frame_capped_not_spiking() {
+        // iterate-3AJ: the regression. A heavy non-presenting frame retires
+        // ~10M instructions; the OLD ticker fired ~66 vsyncs (10M/150k) in
+        // that single frame, jumping the guest vsync counter 0→66 and
+        // skipping the fade-in. The present-anchored ticker caps the dry
+        // window at DRY_FALLBACK_CAP.
+        let mut s = InterruptState::default();
+        s.set_callback(0x1000, 0xAB);
+        // Enter anchored mode via one present.
+        let mut instr: u64 = VSYNC_INSTR_PERIOD;
+        assert!(s.tick_vsync_instr(instr, 1));
+        while s.take_next().is_some() {}
+        // Simulate a 10M-instruction frame with NO new present, ticked in
+        // chunks (as coord_pre_round would). Count fallback vsyncs queued.
+        let mut fallback = 0usize;
+        for _ in 0..100 {
+            instr += 100_000; // 100 chunks × 100k = 10M instructions
+            if s.tick_vsync_instr(instr, 1) {
+                while s.take_next().is_some() {
+                    fallback += 1;
+                }
+            }
+        }
+        assert_eq!(
+            fallback, DRY_FALLBACK_CAP as usize,
+            "a heavy dry frame must cap fallback vsyncs at DRY_FALLBACK_CAP, \
+             not fire ~66"
+        );
+    }
+
     #[test]
     fn tick_vsync_wallclock_first_call_sets_anchor() {
         // First call seeds the anchor and never fires. KRNBUG-D08:

crates/xenia-kernel/src/path.rs

@@ -13,7 +13,7 @@ use xenia_memory::{GuestMemory, MemoryAccess};
 ///   u16 Length
 ///   u16 MaximumLength
 ///   u32 Buffer (guest pointer)
-fn read_ansi_string(mem: &GuestMemory, ptr: u32) -> Option<String> {
+pub fn read_ansi_string(mem: &GuestMemory, ptr: u32) -> Option<String> {
     if ptr == 0 {
         return None;
     }

crates/xenia-kernel/src/state.rs

@@ -219,6 +219,17 @@ pub struct KernelState {
     /// only). Used by `xex_get_procedure_address` to resolve ordinals back
     /// to callable thunks.
     thunks_by_ordinal: HashMap<(ModuleId, u16), u32>,
+
+    /// Perf (Tier-A #4): inclusive [min, max] guest-address band that
+    /// contains every registered import thunk. Import thunks sit in a
+    /// small contiguous region of the XEX; almost every executing PC is
+    /// ordinary guest code OUTSIDE this band. The per-slot-visit prologue
+    /// looks up `thunk_map.get(&pc)` (a `HashMap<u32,…>` → `hash_one` per
+    /// call, ~3.2M visits boot-to-splash). Range-rejecting against this
+    /// band first turns the common (non-thunk) case into a pair of integer
+    /// compares and skips the hash entirely. `None` until the first thunk
+    /// is registered (no band → reject everything, matching an empty map).
+    thunk_addr_band: Option<(u32, u32)>,
     /// First-Pixels diagnostic latch. Set the first time
     /// `RtlRaiseException` fires with code `0xE06D7363` (MSVC C++ throw)
     /// so the deep stack-walk + `runtime_error` decode in
@@ -374,6 +385,15 @@ pub struct KernelState {
     /// block every round from the deterministic `global_clock` via
     /// [`Self::update_timestamp_bundle`].
     pub timestamp_bundle_addr: u32,
+
+    /// Perf (Tier-B #5) throttle state for [`Self::update_timestamp_bundle`].
+    /// Holds the `clock` value at which the bundle was last actually written;
+    /// `u64::MAX` is the "never written" sentinel (forces the first write).
+    /// `AtomicU64` (not `Cell`) so the `&self` update path stays `Sync` for
+    /// the parallel `Arc<Mutex<KernelState>>` usage. Only ever advanced
+    /// forward under the kernel lock, so `Relaxed` ordering is sufficient and
+    /// the sequence is deterministic.
+    timestamp_bundle_last_clock: std::sync::atomic::AtomicU64,
 }
 
 /// ITERATE-2C Phase D — one queued auto-signal. `deadline_cycle` is
@@ -439,6 +459,7 @@ impl KernelState {
             audit: HandleAudit::default(),
             reservations,
             thunks_by_ordinal: HashMap::new(),
+            thunk_addr_band: None,
             cxx_throw_logged: false,
             ring_base: 0,
             ring_size_dwords: 0,
@@ -465,6 +486,7 @@ impl KernelState {
             last_cycle_hint: 0,
             silph_autosignal_diag_logged: false,
             timestamp_bundle_addr: 0,
+            timestamp_bundle_last_clock: std::sync::atomic::AtomicU64::new(u64::MAX),
         };
         crate::exports::register_exports(&mut state);
         crate::xam::register_exports(&mut state);
@@ -584,6 +606,25 @@ impl KernelState {
     /// emits each ordinal once per module).
     pub fn register_thunk(&mut self, module: ModuleId, ordinal: u16, address: u32) {
         self.thunks_by_ordinal.insert((module, ordinal), address);
+        // Widen the thunk address band (Tier-A #4) so the hot prologue can
+        // range-reject non-thunk PCs before hashing the thunk map.
+        self.thunk_addr_band = Some(match self.thunk_addr_band {
+            Some((lo, hi)) => (lo.min(address), hi.max(address)),
+            None => (address, address),
+        });
+    }
+
+    /// Perf (Tier-A #4). Cheap pre-filter for the per-slot-visit import-thunk
+    /// dispatch: `false` guarantees `pc` is NOT a registered thunk (so the
+    /// caller can skip the `thunk_map.get(&pc)` hash). `true` means `pc` lies
+    /// within the registered thunk address band and the map must be consulted
+    /// for an exact match. Conservative — never a false negative.
+    #[inline]
+    pub fn pc_in_thunk_band(&self, pc: u32) -> bool {
+        match self.thunk_addr_band {
+            Some((lo, hi)) => pc >= lo && pc <= hi,
+            None => false,
+        }
     }
 
     /// Resolve a `(module, ordinal)` to its registered thunk address.
@@ -919,6 +960,31 @@ impl KernelState {
             return;
         }
         const INSTRUCTIONS_PER_MS: u64 = 10_000;
+        // Perf (Tier-B #5): the bundle is updated once per scheduler round
+        // (~every 7 retired instructions), but the four guest BE memory
+        // writes are ~8.6% of boot-to-splash. `clock` is the retired-
+        // instruction count, so consecutive rounds rewrite essentially the
+        // same staircase. Throttle to a 0.25 ms quantum: only re-write when
+        // `clock` advanced by >= INSTRUCTIONS_PER_MS/4 (2500 units) since the
+        // last write. This keeps `tick_count` (ms, changes every 10_000
+        // units) ALWAYS fresh and `interrupt_time`/`system_time` monotone at
+        // 0.25 ms granularity — finer than any guest deadline math needs
+        // (`parse_timeout` works in whole ms; the hub gate is `+66 ms`). The
+        // fade-in (3AH-proven vsync-counter driven, NOT this bundle) is
+        // untouched. Throttle threshold is well below 1 ms so no guest-
+        // visible ms boundary is ever skipped.
+        const BUNDLE_QUANTUM: u64 = INSTRUCTIONS_PER_MS / 4; // 2500 units = 0.25 ms
+        {
+            use std::sync::atomic::Ordering;
+            let last = self.timestamp_bundle_last_clock.load(Ordering::Relaxed);
+            // Always allow the first write (last == u64::MAX sentinel) and any
+            // write that crosses the quantum. Never go backwards.
+            if last != u64::MAX && clock < last.saturating_add(BUNDLE_QUANTUM) {
+                return;
+            }
+            self.timestamp_bundle_last_clock
+                .store(clock, Ordering::Relaxed);
+        }
         // FILETIME epoch base (~2021) so `system_time` is a plausible
         // absolute wall-clock; matches the constant used by
         // `ke_query_system_time`. interrupt_time is "since boot" so it
@@ -1042,6 +1108,24 @@ impl KernelState {
         }
     }
 
+    /// Perf gate (Tier-A quick-win #3). `true` iff any of the four
+    /// per-slot-visit diagnostic probe registries
+    /// (`ctor_probe_pcs` / `branch_probe_pcs` / `audit_pc_probe_pcs`
+    /// / `lr_trace_pcs`) holds at least one PC. The common headless
+    /// run leaves all four empty, so the prologue can skip the four
+    /// `fire_*_if_match` calls entirely with this single predicted
+    /// branch — avoiding 4× call overhead per slot-visit (~3.2M
+    /// visits over boot-to-splash) when no probe is configured.
+    /// Purely a fast-path guard; each `fire_*` still re-checks its own
+    /// set, so behaviour is identical whether or not the caller gates.
+    #[inline]
+    pub fn any_probe_active(&self) -> bool {
+        !self.ctor_probe_pcs.is_empty()
+            || !self.branch_probe_pcs.is_empty()
+            || !self.audit_pc_probe_pcs.is_empty()
+            || !self.lr_trace_pcs.is_empty()
+    }
+
     /// Diagnostic. If the live PC for HW slot `hw_id` is in
     /// `self.ctor_probe_pcs`, emit a single `CTOR-PROBE` line with
     /// the current cycle, tid, hw_id, sp, r3, lr, plus an 8-frame

crates/xenia-kernel/src/ui_bridge.rs

@@ -14,6 +14,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU64};
 
+use xenia_gpu::draw_capture::DrawCapture;
 use xenia_gpu::texture_cache::TextureKey;
 use xenia_gpu::xenos_constants::XenosConstantsBlock;
 use xenia_hid::GamepadState;
@@ -133,6 +134,14 @@ pub struct UiBridge {
     /// reverts to its magenta stub.
     pub publish_texture:
         Arc<dyn Fn(Option<(TextureKey, Vec<u8>)>) + Send + Sync>,
+    /// iterate-3O real-render slice: at each `VdSwap`, the kernel hands the
+    /// UI the per-draw geometry captured this frame (one [`DrawCapture`] per
+    /// `PM4_DRAW_INDX*`), including the real guest vertex window. The UI
+    /// replays them through the Xenos wgpu pipeline so the splash renders its
+    /// actual geometry instead of synthetic placeholder shapes. Empty in the
+    /// degenerate case (no draws or capture disabled).
+    pub publish_geometry:
+        Arc<dyn Fn(Vec<DrawCapture>) + Send + Sync>,
 }
 
 impl UiBridge {
@@ -182,4 +191,9 @@ impl UiBridge {
     pub fn publish_texture(&self, tex: Option<(TextureKey, Vec<u8>)>) {
         (self.publish_texture)(tex);
     }
+
+    /// Hand this frame's captured per-draw geometry to the UI.
+    pub fn publish_geometry(&self, caps: Vec<DrawCapture>) {
+        (self.publish_geometry)(caps);
+    }
 }

crates/xenia-memory/src/heap.rs

@@ -89,6 +89,14 @@ pub struct GuestMemory {
     mem_watch_addrs: Vec<u32>,
     /// Count of fires observed (for tests / hand-off telemetry).
     mem_watch_count: AtomicU64,
+    /// Monotonic count of MMIO accesses (every scalar load/store that
+    /// resolves to a registered MMIO region bumps this by 1). A pure,
+    /// deterministic function of guest execution — the superblock runner
+    /// samples it before/after each block to detect an MMIO touch and
+    /// end the run there (so MMIO ordering vs other HW threads stays at
+    /// the same fine lockstep granularity as before). Relaxed because the
+    /// lockstep path is single-threaded and only needs monotonicity.
+    mmio_access_count: AtomicU64,
 }
 
 /// Greatest common bit-mask such that `(a & m) == (b & m)` for every bit
@@ -133,9 +141,26 @@ impl GuestMemory {
             writes_total: AtomicU64::new(0),
             mem_watch_addrs: Vec::new(),
             mem_watch_count: AtomicU64::new(0),
+            mmio_access_count: AtomicU64::new(0),
         })
     }
 
+    /// Monotonic count of MMIO accesses since boot. Used by the superblock
+    /// runner to detect that a just-executed block touched MMIO (so it can
+    /// end the superblock there and keep MMIO ordering at lockstep
+    /// granularity). Deterministic function of guest execution.
+    #[inline]
+    pub fn mmio_access_count(&self) -> u64 {
+        self.mmio_access_count
+            .load(std::sync::atomic::Ordering::Relaxed)
+    }
+
+    #[inline]
+    fn bump_mmio_access(&self) {
+        self.mmio_access_count
+            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+    }
+
     /// Current version watermark for the page containing `addr`. Bumped by
     /// any write through `write_u8/16/32/64`. Not affected by MMIO writes
     /// (those don't touch the backing texture memory).
@@ -357,7 +382,8 @@ impl GuestMemory {
     /// from `GuestMemory` without a wider plumbing change).
     pub fn write_bulk(&self, addr: u32, buf: &[u8]) {
         let len = buf.len() as u32;
-        let old_lane = self.capture_mem_watch_old(addr, len);
+        let watch = self.has_mem_watch();
+        let old_lane = if watch { self.capture_mem_watch_old(addr, len) } else { None };
         let ptr = self.translate_virtual_mut(addr);
         unsafe {
             std::ptr::copy_nonoverlapping(buf.as_ptr(), ptr, buf.len());
@@ -374,7 +400,7 @@ impl GuestMemory {
             // the page works.
             self.bump_page_version(page * PAGE_SIZE);
         }
-        self.check_mem_watch(addr, len, old_lane);
+        if watch { self.check_mem_watch(addr, len, old_lane); }
     }
 
     /// Check if a guest address has been allocated/committed. Acquire load
@@ -487,6 +513,7 @@ impl MemoryAccess for GuestMemory {
         // MMIO dispatch must come first — a byte read at an MMIO-mapped
         // address should invoke the callback, not the backing memory.
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             return (mmio.read_callback)(addr) as u8;
         }
         if !self.is_mapped(addr) { return 0; }
@@ -497,6 +524,7 @@ impl MemoryAccess for GuestMemory {
     #[inline]
     fn read_u16(&self, addr: u32) -> u16 {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.read_callback)(addr) as u16
         } else if !self.is_mapped(addr) {
             0
@@ -509,6 +537,7 @@ impl MemoryAccess for GuestMemory {
     #[inline]
     fn read_u32(&self, addr: u32) -> u32 {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.read_callback)(addr)
         } else if !self.is_mapped(addr) {
             0
@@ -521,6 +550,7 @@ impl MemoryAccess for GuestMemory {
     #[inline]
     fn read_u64(&self, addr: u32) -> u64 {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             let hi = (mmio.read_callback)(addr) as u64;
             let lo = (mmio.read_callback)(addr.wrapping_add(4)) as u64;
             (hi << 32) | lo
@@ -536,23 +566,31 @@ impl MemoryAccess for GuestMemory {
         // MMIO dispatch first — a byte write at an MMIO-mapped address
         // must invoke the callback, not the backing memory.
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.write_callback)(addr, val as u32);
             return;
         }
         if !self.is_mapped(addr) { return; }
-        let old_lane = self.capture_mem_watch_old(addr, 1);
+        // Perf (Tier-A #1): the mem-watch capture/report pair are out-of-line
+        // calls; on the common (no-watch) path each was a real call that
+        // immediately returned. Gate both behind one predicted branch so the
+        // hot store does no call work unless a watch is actually armed.
+        let watch = self.has_mem_watch();
+        let old_lane = if watch { self.capture_mem_watch_old(addr, 1) } else { None };
         let ptr = self.translate_virtual_mut(addr);
         unsafe { *ptr = val };
         self.bump_page_version(addr);
-        self.check_mem_watch(addr, 1, old_lane);
+        if watch { self.check_mem_watch(addr, 1, old_lane); }
     }
 
     fn write_u16(&self, addr: u32, val: u16) {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.write_callback)(addr, val as u32);
         } else if !self.is_mapped(addr) {
         } else {
-            let old_lane = self.capture_mem_watch_old(addr, 2);
+            let watch = self.has_mem_watch();
+            let old_lane = if watch { self.capture_mem_watch_old(addr, 2) } else { None };
             let ptr = self.translate_virtual_mut(addr);
             unsafe {
                 std::ptr::copy_nonoverlapping(val.to_be_bytes().as_ptr(), ptr, 2);
@@ -564,16 +602,18 @@ impl MemoryAccess for GuestMemory {
             if (addr & 0xFFF) >= (PAGE_SIZE - 1) {
                 self.bump_page_version(addr.wrapping_add(1));
             }
-            self.check_mem_watch(addr, 2, old_lane);
+            if watch { self.check_mem_watch(addr, 2, old_lane); }
         }
     }
 
     fn write_u32(&self, addr: u32, val: u32) {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.write_callback)(addr, val);
         } else if !self.is_mapped(addr) {
         } else {
-            let old_lane = self.capture_mem_watch_old(addr, 4);
+            let watch = self.has_mem_watch();
+            let old_lane = if watch { self.capture_mem_watch_old(addr, 4) } else { None };
             let ptr = self.translate_virtual_mut(addr);
             unsafe {
                 std::ptr::copy_nonoverlapping(val.to_be_bytes().as_ptr(), ptr, 4);
@@ -582,17 +622,19 @@ impl MemoryAccess for GuestMemory {
             if (addr & 0xFFF) >= (PAGE_SIZE - 3) {
                 self.bump_page_version(addr.wrapping_add(3));
             }
-            self.check_mem_watch(addr, 4, old_lane);
+            if watch { self.check_mem_watch(addr, 4, old_lane); }
         }
     }
 
     fn write_u64(&self, addr: u32, val: u64) {
         if let Some(mmio) = self.find_mmio(addr) {
+            self.bump_mmio_access();
             (mmio.write_callback)(addr, (val >> 32) as u32);
             (mmio.write_callback)(addr.wrapping_add(4), val as u32);
         } else if !self.is_mapped(addr) {
         } else {
-            let old_lane = self.capture_mem_watch_old(addr, 8);
+            let watch = self.has_mem_watch();
+            let old_lane = if watch { self.capture_mem_watch_old(addr, 8) } else { None };
             let ptr = self.translate_virtual_mut(addr);
             unsafe {
                 std::ptr::copy_nonoverlapping(val.to_be_bytes().as_ptr(), ptr, 8);
@@ -601,7 +643,7 @@ impl MemoryAccess for GuestMemory {
             if (addr & 0xFFF) >= (PAGE_SIZE - 7) {
                 self.bump_page_version(addr.wrapping_add(7));
             }
-            self.check_mem_watch(addr, 8, old_lane);
+            if watch { self.check_mem_watch(addr, 8, old_lane); }
         }
     }
 

crates/xenia-ui/src/app.rs

@@ -181,10 +181,11 @@ impl App {
         y += line_h;
         let (fbw, fbh) = rs.frontbuffer_size();
         let render_line = format!(
-            "Render: xdispatch: xlated={:>5} interp={:>5}  xlated-pipelines={:>3}  tex-cache={:>3}  fb={}x{}",
+            "Render: xdispatch: xlated={:>5} interp={:>5}  xlated-pipelines={:>3}  real-geo={:>5}  tex-cache={:>3}  fb={}x{}",
             rs.xenos_dispatches_translator,
             rs.xenos_dispatches_interpreter,
             rs.translated_pipeline_count(),
+            rs.real_geometry_draws(),
             rs.host_texture_count(),
             fbw,
             fbh,
@@ -368,53 +369,28 @@ impl ApplicationHandler<SwapEvent> for App {
                             .map(|s| s.frame_index)
                             .unwrap_or(0);
                         if frame_idx != self.last_xenos_swap_frame {
-                            rs.clear_frontbuffer([0.04, 0.04, 0.06, 1.0]);
+                            // iterate-3AE: clear to BLACK, matching canary's
+                            // splash background. The old navy `[0.04,0.04,0.06]`
+                            // was an iterate-3S debug placeholder never matched
+                            // to the guest. The splash background-fill draw is a
+                            // full-screen Xbox-360 RectangleList (3 verts → a HW
+                            // rectangle covering the whole screen); the UI replay
+                            // draws it as a single triangle (the 4th implied
+                            // corner isn't synthesized), so only the diagonal
+                            // half is covered. With a navy clear the uncovered
+                            // half showed a navy diagonal in the brief
+                            // pre/inter-logo transition frames (where that fill
+                            // is the only coverage). Canary's background there is
+                            // black, and the guest's fill itself resolves to
+                            // black, so a black clear makes the uncovered half
+                            // match — the transition is uniformly black like the
+                            // oracle. (Full RectangleList→rectangle expansion is
+                            // the deeper fix and a separate follow-up; under a
+                            // black clear the half-coverage is invisible.)
+                            rs.clear_frontbuffer([0.0, 0.0, 0.0, 1.0]);
                             self.last_xenos_swap_frame = frame_idx;
                         }
                         let delta = (draws_total - already) as u32;
-                        let (verts_hint, prim_kind, vs_key, ps_key) = self
-                            .last_swap_info
-                            .map(|s| {
-                                (
-                                    s.last_draw_vertex_count.max(3),
-                                    s.last_draw_prim,
-                                    s.vs_blob_key,
-                                    s.ps_blob_key,
-                                )
-                            })
-                            .unwrap_or((3, 4, 0, 0));
-                        // Look up blobs + constants from the bridge and
-                        // pack into the WGSL-interpreter layout. Empty
-                        // slices produce zero-clause packed buffers — the
-                        // WGSL walker short-circuits and the placeholder
-                        // export path still renders.
-                        let raw_vs: Vec<u32> = self
-                            .handles
-                            .shader_blobs
-                            .lock()
-                            .ok()
-                            .and_then(|g| g.get(&vs_key).cloned())
-                            .unwrap_or_default();
-                        let raw_ps: Vec<u32> = self
-                            .handles
-                            .shader_blobs
-                            .lock()
-                            .ok()
-                            .and_then(|g| g.get(&ps_key).cloned())
-                            .unwrap_or_default();
-                        let parsed_vs = xenia_gpu::ucode::parse_shader(&raw_vs);
-                        let parsed_ps = xenia_gpu::ucode::parse_shader(&raw_ps);
-                        // First time we see a blob key, run the static
-                        // metrics analyzer. Keyed on (stage_tag, blob_key)
-                        // because the guest can reuse a key across stages.
-                        if self.seen_shader_blobs.insert((0u8, vs_key)) {
-                            xenia_gpu::shader_metrics::emit_for(&parsed_vs, "vs");
-                        }
-                        if self.seen_shader_blobs.insert((1u8, ps_key)) {
-                            xenia_gpu::shader_metrics::emit_for(&parsed_ps, "ps");
-                        }
-                        let vs_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_vs);
-                        let ps_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_ps);
                         let constants = self
                             .handles
                             .xenos_constants
@@ -431,6 +407,58 @@ impl ApplicationHandler<SwapEvent> for App {
                             .ok()
                             .and_then(|g| g.clone());
                         rs.bind_primary_texture(tex_payload);
+
+                        // iterate-3O real-render slice: prefer replaying the
+                        // *real* captured guest geometry. The kernel publishes
+                        // one `DrawCapture` per `PM4_DRAW_INDX*` this frame
+                        // (real vertices + prim type + shader keys). Fall back
+                        // to the legacy synthetic dispatch only when no capture
+                        // is available (e.g. capture disabled), so we never
+                        // regress to a blank screen.
+                        let captures: Vec<xenia_gpu::draw_capture::DrawCapture> = self
+                            .handles
+                            .geometry
+                            .lock()
+                            .map(|g| g.clone())
+                            .unwrap_or_default();
+                        let blobs: std::collections::HashMap<u32, Vec<u32>> = self
+                            .handles
+                            .shader_blobs
+                            .lock()
+                            .map(|g| g.clone())
+                            .unwrap_or_default();
+                        if !captures.is_empty() {
+                            rs.dispatch_xenos_captures(
+                                &captures,
+                                &blobs,
+                                &constants,
+                                &mut self.seen_shader_blobs,
+                            );
+                        } else {
+                            // Legacy synthetic-geometry fallback (placeholder).
+                            let (verts_hint, prim_kind, vs_key, ps_key) = self
+                                .last_swap_info
+                                .map(|s| {
+                                    (
+                                        s.last_draw_vertex_count.max(3),
+                                        s.last_draw_prim,
+                                        s.vs_blob_key,
+                                        s.ps_blob_key,
+                                    )
+                                })
+                                .unwrap_or((3, 4, 0, 0));
+                            let raw_vs = blobs.get(&vs_key).cloned().unwrap_or_default();
+                            let raw_ps = blobs.get(&ps_key).cloned().unwrap_or_default();
+                            let parsed_vs = xenia_gpu::ucode::parse_shader(&raw_vs);
+                            let parsed_ps = xenia_gpu::ucode::parse_shader(&raw_ps);
+                            if self.seen_shader_blobs.insert((0u8, vs_key)) {
+                                xenia_gpu::shader_metrics::emit_for(&parsed_vs, "vs");
+                            }
+                            if self.seen_shader_blobs.insert((1u8, ps_key)) {
+                                xenia_gpu::shader_metrics::emit_for(&parsed_ps, "ps");
+                            }
+                            let vs_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_vs);
+                            let ps_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_ps);
                             rs.dispatch_xenos_draws(
                                 already,
                                 delta,
@@ -445,6 +473,7 @@ impl ApplicationHandler<SwapEvent> for App {
                                 &constants,
                             );
                         }
+                    }
                 } else {
                     Self::ingest_frontbuffer(
                         &self.handles,

crates/xenia-ui/src/bridge.rs

@@ -18,6 +18,7 @@ use std::sync::Mutex;
 
 use crossbeam_utils::atomic::AtomicCell;
 use winit::event_loop::EventLoopProxy;
+use xenia_gpu::draw_capture::DrawCapture;
 use xenia_gpu::texture_cache::TextureKey;
 use xenia_gpu::xenos_constants::XenosConstantsBlock;
 use xenia_hid::GamepadState;
@@ -66,6 +67,10 @@ pub struct UiHandles {
     /// fetch-constant slot 0 into linear bytes that the UI should
     /// upload into the host cache and bind at `@group(1) @binding(0)`.
     pub primary_texture: Arc<Mutex<Option<(TextureKey, Vec<u8>)>>>,
+    /// iterate-3O: the most recent frame's captured per-draw geometry. The
+    /// redraw path drains this to replay real guest draws. Replaced wholesale
+    /// each `VdSwap`.
+    pub geometry: Arc<Mutex<Vec<DrawCapture>>>,
 }
 
 /// Swap event posted by the CPU-side `VdSwap` handler via
@@ -89,6 +94,7 @@ pub fn build(proxy: EventLoopProxy<SwapEvent>) -> (UiHandles, UiBridge) {
     let xenos_constants = Arc::new(Mutex::new(XenosConstantsBlock::default()));
     let primary_texture: Arc<Mutex<Option<(TextureKey, Vec<u8>)>>> =
         Arc::new(Mutex::new(None));
+    let geometry: Arc<Mutex<Vec<DrawCapture>>> = Arc::new(Mutex::new(Vec::new()));
 
     let kernel_bridge = UiBridge {
         gamepad: {
@@ -144,6 +150,14 @@ pub fn build(proxy: EventLoopProxy<SwapEvent>) -> (UiHandles, UiBridge) {
                 }
             })
         },
+        publish_geometry: {
+            let geo = Arc::clone(&geometry);
+            Arc::new(move |caps| {
+                if let Ok(mut lock) = geo.lock() {
+                    *lock = caps;
+                }
+            })
+        },
     };
 
     let handles = UiHandles {
@@ -155,6 +169,7 @@ pub fn build(proxy: EventLoopProxy<SwapEvent>) -> (UiHandles, UiBridge) {
         shader_blobs,
         xenos_constants,
         primary_texture,
+        geometry,
     };
     (handles, kernel_bridge)
 }

crates/xenia-ui/src/render.rs

@@ -84,6 +84,9 @@ pub struct RenderState {
     /// the shader, or (c) we're running the slow interpreter path.
     pub xenos_dispatches_translator: u64,
     pub xenos_dispatches_interpreter: u64,
+    /// iterate-3O: running total of replayed draws that carried a real guest
+    /// vertex window (vs. the procedural fallback). Surfaced on the HUD.
+    real_geometry_draws: u64,
     /// One-shot latch so we emit a tracing::info! on the **first** real
     /// draw dispatch rather than spamming every frame. Pairs with the
     /// "first translator compile" latch below.
@@ -447,6 +450,7 @@ impl RenderState {
             fallback_rgb: [0.06, 0.06, 0.09],
             xenos_pipeline,
             xenos_draws_rendered: 0,
+            real_geometry_draws: 0,
             xenos_dispatches_translator: 0,
             xenos_dispatches_interpreter: 0,
             first_dispatch_logged: false,
@@ -657,26 +661,39 @@ impl RenderState {
                 draw_index: idx,
                 vertex_count: vertex_count_hint.max(3),
                 prim_kind,
+                // Synthetic fallback path: no real vertex window.
+                vertex_base_dwords: 0,
+                // No real geometry → no NDC transform (procedural positions are
+                // already in clip space).
+                ndc_scale: [0.0, 0.0],
+                ndc_offset: [0.0, 0.0],
             };
+            // Synthetic visualizer path (legacy): no captured render state, so
+            // use the opaque default.
+            let rstate = crate::xenos_pipeline::RenderState::OPAQUE;
             if use_translated
-                && let Some(p) = self.xenos_pipeline.translated_pipeline(vs_key, ps_key) {
-                    self.xenos_pipeline.render_one_with_pipeline(
+                && self.xenos_pipeline.render_one_translated(
+                    &self.device,
                     &self.queue,
                     &mut encoder,
                     &self.frontbuffer_view,
                     req,
-                        p,
-                    );
-                    metrics::counter!("gpu.shader.use", "path" => "translator")
-                        .increment(1);
+                    vs_key,
+                    ps_key,
+                    rstate,
+                )
+            {
+                metrics::counter!("gpu.shader.use", "path" => "translator").increment(1);
                 served_translator += 1;
                 continue;
             }
             self.xenos_pipeline.render_one(
+                &self.device,
                 &self.queue,
                 &mut encoder,
                 &self.frontbuffer_view,
                 req,
+                rstate,
             );
             metrics::counter!("gpu.shader.use", "path" => "interpreter").increment(1);
             served_interpreter += 1;
@@ -707,12 +724,201 @@ impl RenderState {
         }
     }
 
+    /// iterate-3O real-render slice: replay a batch of *real* captured guest
+    /// draws. Unlike [`dispatch_xenos_draws`] (synthetic placeholder geometry),
+    /// each [`DrawCapture`] carries the actual guest vertex window, primitive
+    /// type, host vertex count, and the real (vs, ps) keys. Per capture we:
+    ///   1. upload the captured guest vertex bytes into `vertex_buffer` (b4),
+    ///   2. upload the matching VS/PS microcode + per-frame constants,
+    ///   3. render through the translated (P7) pipeline if it compiled, else
+    ///      the interpreter — with `vertex_base_dwords` set so the shader
+    ///      rebases its absolute fetch address into the uploaded window.
+    ///
+    /// Returns the number of captures that had a real vertex window (vs. the
+    /// procedural fallback), for HUD reporting. `shader_blobs` / `constants`
+    /// come from the bridge; `seen` records which blobs have had static
+    /// metrics emitted (one-shot per blob, matching the legacy path).
+    pub fn dispatch_xenos_captures(
+        &mut self,
+        captures: &[xenia_gpu::draw_capture::DrawCapture],
+        shader_blobs: &std::collections::HashMap<u32, Vec<u32>>,
+        constants: &xenia_gpu::xenos_constants::XenosConstantsBlock,
+        seen: &mut std::collections::HashSet<(u8, u32)>,
+    ) -> u32 {
+        if captures.is_empty() {
+            return 0;
+        }
+        let mut real_count = 0u32;
+        // iterate-3X (GPUBUG-111): each captured draw uploads its OWN vertex
+        // window + per-draw constants + shader via `queue.write_buffer`. In
+        // wgpu all `write_buffer` calls staged before a single `queue.submit`
+        // are applied *before any* command in that submit executes — so a single
+        // encoder for the whole batch made every draw read only the LAST draw's
+        // vertex buffer / uniforms (the splash logo quad sampled the fullscreen
+        // background quad's vertices → nothing rendered where the logo was).
+        // Submit ONE encoder PER draw so each draw's writes land before its own
+        // pass. The frontbuffer uses `LoadOp::Load`, so per-draw submits still
+        // composite over each other exactly like before.
+        for cap in captures {
+            // iterate-3T: bind this draw's REAL decoded texture (keyed off the
+            // active PS's tfetch slot, attached in `gpu_system`) so the textured
+            // logo samples the artwork. `None` reverts to the magenta stub for
+            // flat draws. Each `set_texture_view` rebuilds the tex bind group;
+            // the subsequent `render_one*` reads it, so per-draw binding works
+            // even though all draws share one encoder.
+            {
+                let Self {
+                    device,
+                    queue,
+                    xenos_pipeline,
+                    host_texture_cache,
+                    ..
+                } = self;
+                match cap.textures.first() {
+                    Some((key, version, bytes)) => {
+                        // iterate-3AD: use the decoder's real content `version`
+                        // (from `span_max_version`) so the host cache re-uploads
+                        // when the guest fills MORE of an evolving atlas. The
+                        // publisher and the 2nd splash logo share one K8888
+                        // surface (base 0x4dbee000); the 2nd logo's texels land
+                        // AFTER the first upload. With the old hardcoded
+                        // `version_when_uploaded = 1`, the same `TextureKey`
+                        // never re-uploaded, so the 2nd logo sampled its (then
+                        // still-zero) atlas region as black. The real version
+                        // increases as the guest writes, triggering re-upload.
+                        let cached = xenia_gpu::texture_cache::CachedTexture {
+                            key: *key,
+                            version_when_uploaded: *version,
+                            bytes: bytes.clone(),
+                        };
+                        host_texture_cache.upload(device, queue, &cached);
+                        if let Some(view) = host_texture_cache.view_for(key) {
+                            xenos_pipeline.set_texture_view(device, Some(view));
+                        }
+                    }
+                    None => xenos_pipeline.set_texture_view(device, None),
+                }
+            }
+            let raw_vs = shader_blobs.get(&cap.vs_key).cloned().unwrap_or_default();
+            let raw_ps = shader_blobs.get(&cap.ps_key).cloned().unwrap_or_default();
+            let parsed_vs = xenia_gpu::ucode::parse_shader(&raw_vs);
+            let parsed_ps = xenia_gpu::ucode::parse_shader(&raw_ps);
+            if seen.insert((0u8, cap.vs_key)) {
+                xenia_gpu::shader_metrics::emit_for(&parsed_vs, "vs");
+            }
+            if seen.insert((1u8, cap.ps_key)) {
+                xenia_gpu::shader_metrics::emit_for(&parsed_ps, "ps");
+            }
+            let vs_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_vs);
+            let ps_packed = xenia_gpu::ucode::pack_for_wgsl(&parsed_ps);
+            // Upload this draw's shader + constants + real vertex window.
+            self.xenos_pipeline.upload_shader_and_constants(
+                &self.queue,
+                &vs_packed,
+                &ps_packed,
+                constants,
+            );
+            if cap.has_real_vertices && !cap.vertex_dwords.is_empty() {
+                self.xenos_pipeline
+                    .upload_vertex_data(&self.queue, &cap.vertex_dwords);
+                real_count += 1;
+            }
+            let use_translated = cap.vs_key != 0
+                && cap.ps_key != 0
+                && ensure_translated_pipeline(
+                    &mut self.xenos_pipeline,
+                    &self.device,
+                    cap.vs_key,
+                    cap.ps_key,
+                    &parsed_vs,
+                    &parsed_ps,
+                );
+            let base = if cap.has_real_vertices {
+                cap.window_base_dwords
+            } else {
+                0
+            };
+            let req = DrawRequest {
+                draw_index: cap.draw_index,
+                vertex_count: cap.host_vertex_count.max(3),
+                prim_kind: cap.prim_code,
+                vertex_base_dwords: base,
+                // iterate-3S: apply the per-draw guest viewport → host NDC
+                // transform only when we have real geometry (otherwise the
+                // procedural fallback already emits clip-space positions).
+                ndc_scale: if cap.has_real_vertices { cap.ndc_scale } else { [0.0, 0.0] },
+                ndc_offset: if cap.has_real_vertices { cap.ndc_offset } else { [0.0, 0.0] },
+            };
+            // iterate-3Y: replay this draw's real color/blend/write-mask state
+            // (captured from `RB_BLENDCONTROL0` / `RB_COLOR_MASK`) so overlays
+            // composite the way the guest intends instead of opaquely
+            // overwriting the logo.
+            let rstate = crate::xenos_pipeline::RenderState {
+                blend_control: cap.blend_control,
+                color_mask: cap.color_mask,
+            };
+            let mut encoder = self
+                .device
+                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
+                    label: Some("xenos capture replay (per-draw)"),
+                });
+            let served_translated = use_translated
+                && self.xenos_pipeline.render_one_translated(
+                    &self.device,
+                    &self.queue,
+                    &mut encoder,
+                    &self.frontbuffer_view,
+                    req,
+                    cap.vs_key,
+                    cap.ps_key,
+                    rstate,
+                );
+            if served_translated {
+                self.xenos_dispatches_translator =
+                    self.xenos_dispatches_translator.saturating_add(1);
+            } else {
+                self.xenos_pipeline.render_one(
+                    &self.device,
+                    &self.queue,
+                    &mut encoder,
+                    &self.frontbuffer_view,
+                    req,
+                    rstate,
+                );
+                self.xenos_dispatches_interpreter =
+                    self.xenos_dispatches_interpreter.saturating_add(1);
+            }
+            self.queue.submit(std::iter::once(encoder.finish()));
+        }
+        self.xenos_draws_rendered = self
+            .xenos_draws_rendered
+            .saturating_add(captures.len() as u64);
+        self.real_geometry_draws = self
+            .real_geometry_draws
+            .saturating_add(real_count as u64);
+        if !self.first_dispatch_logged {
+            self.first_dispatch_logged = true;
+            tracing::info!(
+                captures = captures.len(),
+                real_vertex_draws = real_count,
+                "first Xenos capture batch replayed (real geometry)"
+            );
+        }
+        real_count
+    }
+
     /// Count of distinct translator pipelines compiled so far. Surfaced
     /// on the HUD as `xlated=N` to make "is P7 working?" observable.
     pub fn translated_pipeline_count(&self) -> usize {
         self.xenos_pipeline.translated_pipeline_count()
     }
 
+    /// Running count of captured draws that carried a real vertex window
+    /// (surfaced on the HUD). Updated by [`dispatch_xenos_captures`].
+    pub fn real_geometry_draws(&self) -> u64 {
+        self.real_geometry_draws
+    }
+
     /// Clear the frontbuffer to `[r,g,b,a]` in linear space. Matches the
     /// fallback clear the outer swapchain render does so the two stages
     /// agree on "no draws yet = dark navy".

crates/xenia-ui/src/xenos_pipeline.rs

@@ -36,7 +36,142 @@ struct DrawConstants {
     draw_index: u32,
     vertex_count: u32,
     prim_kind: u32,
-    _pad: u32,
+    /// iterate-3O: guest dword base of the uploaded `vertex_buffer` window.
+    /// The WGSL subtracts this from the absolute vertex-fetch address.
+    vertex_base_dwords: u32,
+    /// iterate-3S: guest→host NDC XY transform (mirrors canary
+    /// `GetHostViewportInfo`). `clip.xy = pos.xy * ndc_scale + ndc_offset*pos.w`.
+    /// Y is pre-flipped for wgpu. 16 bytes so the block stays 16-byte aligned.
+    ndc_scale: [f32; 2],
+    ndc_offset: [f32; 2],
+}
+
+/// iterate-3Y: the per-draw host color/blend/write-mask render state, decoded
+/// from the guest registers (`RB_BLENDCONTROL0` / `RB_COLOR_MASK`). Used both
+/// as part of the pipeline-cache key and to build the `wgpu::ColorTargetState`.
+/// Mirrors canary's `GetColorBlendStateForRenderTarget` (D3D12
+/// `pipeline_cache.cc`): the factors come straight from `RB_BLENDCONTROL`,
+/// and a zero write-mask forces the no-blend `One,Zero` equation.
+#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
+pub struct RenderState {
+    /// `RB_BLENDCONTROL0` raw value (RT0). `0x00010001` (One,Zero / One,Zero,
+    /// Add) is the opaque case.
+    pub blend_control: u32,
+    /// RT0 nibble of `RB_COLOR_MASK` (bit0=R … bit3=A). 0 = write nothing.
+    pub color_mask: u8,
+}
+
+impl RenderState {
+    /// Fully-opaque, all-channels state (the legacy fixed behaviour). Used for
+    /// procedural/synthetic draws that have no captured guest state.
+    pub const OPAQUE: RenderState = RenderState {
+        blend_control: 0x0001_0001,
+        color_mask: 0xF,
+    };
+
+    /// Map a Xenos `BlendFactor` (5-bit field) to a wgpu `BlendFactor`,
+    /// mirroring canary `kBlendFactorMap` (D3D12 `pipeline_cache.cc:1504`).
+    fn map_factor(f: u32) -> wgpu::BlendFactor {
+        match f {
+            0 => wgpu::BlendFactor::Zero,
+            1 => wgpu::BlendFactor::One,
+            4 => wgpu::BlendFactor::Src,
+            5 => wgpu::BlendFactor::OneMinusSrc,
+            6 => wgpu::BlendFactor::SrcAlpha,
+            7 => wgpu::BlendFactor::OneMinusSrcAlpha,
+            8 => wgpu::BlendFactor::Dst,
+            9 => wgpu::BlendFactor::OneMinusDst,
+            10 => wgpu::BlendFactor::DstAlpha,
+            11 => wgpu::BlendFactor::OneMinusDstAlpha,
+            12 => wgpu::BlendFactor::Constant,
+            13 => wgpu::BlendFactor::OneMinusConstant,
+            14 => wgpu::BlendFactor::Constant,
+            15 => wgpu::BlendFactor::OneMinusConstant,
+            16 => wgpu::BlendFactor::SrcAlphaSaturated,
+            // 2/3 and >16 are undefined on Xenos; canary maps to Zero.
+            _ => wgpu::BlendFactor::Zero,
+        }
+    }
+
+    /// Map a Xenos `BlendFactor` for the *alpha* channel, mirroring canary
+    /// `kBlendFactorAlphaMap` (color-mode factors collapse to alpha).
+    fn map_factor_alpha(f: u32) -> wgpu::BlendFactor {
+        match f {
+            4 => wgpu::BlendFactor::SrcAlpha,
+            5 => wgpu::BlendFactor::OneMinusSrcAlpha,
+            8 => wgpu::BlendFactor::DstAlpha,
+            9 => wgpu::BlendFactor::OneMinusDstAlpha,
+            other => Self::map_factor(other),
+        }
+    }
+
+    fn map_op(o: u32) -> wgpu::BlendOperation {
+        match o {
+            0 => wgpu::BlendOperation::Add,
+            1 => wgpu::BlendOperation::Subtract,
+            2 => wgpu::BlendOperation::Min,
+            3 => wgpu::BlendOperation::Max,
+            4 => wgpu::BlendOperation::ReverseSubtract,
+            _ => wgpu::BlendOperation::Add,
+        }
+    }
+
+    /// Build the `wgpu::ColorTargetState` for this draw.
+    fn color_target(&self, format: wgpu::TextureFormat) -> wgpu::ColorTargetState {
+        let bc = self.blend_control;
+        let color_src = bc & 0x1F;
+        let color_op = (bc >> 5) & 0x7;
+        let color_dst = (bc >> 8) & 0x1F;
+        let alpha_src = (bc >> 16) & 0x1F;
+        let alpha_op = (bc >> 21) & 0x7;
+        let alpha_dst = (bc >> 24) & 0x1F;
+
+        // wgpu requires `blend: None` when nothing would be written; also the
+        // `One,Zero,Add` identity is the opaque case (canary's no-blend), which
+        // we express as `blend: None` so it's a plain overwrite.
+        let is_opaque = color_src == 1
+            && color_dst == 0
+            && color_op == 0
+            && alpha_src == 1
+            && alpha_dst == 0
+            && alpha_op == 0;
+        let blend = if is_opaque {
+            None
+        } else {
+            Some(wgpu::BlendState {
+                color: wgpu::BlendComponent {
+                    src_factor: Self::map_factor(color_src),
+                    dst_factor: Self::map_factor(color_dst),
+                    operation: Self::map_op(color_op),
+                },
+                alpha: wgpu::BlendComponent {
+                    src_factor: Self::map_factor_alpha(alpha_src),
+                    dst_factor: Self::map_factor_alpha(alpha_dst),
+                    operation: Self::map_op(alpha_op),
+                },
+            })
+        };
+
+        let mut write_mask = wgpu::ColorWrites::empty();
+        if self.color_mask & 0x1 != 0 {
+            write_mask |= wgpu::ColorWrites::RED;
+        }
+        if self.color_mask & 0x2 != 0 {
+            write_mask |= wgpu::ColorWrites::GREEN;
+        }
+        if self.color_mask & 0x4 != 0 {
+            write_mask |= wgpu::ColorWrites::BLUE;
+        }
+        if self.color_mask & 0x8 != 0 {
+            write_mask |= wgpu::ColorWrites::ALPHA;
+        }
+
+        wgpu::ColorTargetState {
+            format,
+            blend,
+            write_mask,
+        }
+    }
 }
 
 /// Submitted to [`XenosPipeline::render_one`] to render one captured draw.
@@ -48,6 +183,13 @@ pub struct DrawRequest {
     pub vertex_count: u32,
     /// Xenos primitive-type code; shader may branch on it in P3b+.
     pub prim_kind: u32,
+    /// iterate-3O: guest dword base of the per-draw vertex window uploaded to
+    /// `vertex_buffer` (b4). 0 = no real vertex window (procedural fallback).
+    pub vertex_base_dwords: u32,
+    /// iterate-3S: guest→host NDC XY transform (Y pre-flipped). When all-zero
+    /// the shader leaves the position untransformed (procedural fallback).
+    pub ndc_scale: [f32; 2],
+    pub ndc_offset: [f32; 2],
 }
 
 /// Reasonable upper bound on a single shader blob (dwords). Most Xbox 360
@@ -57,7 +199,16 @@ const UCODE_BUFFER_MAX_DWORDS: u64 = 16 * 1024; // 64 KB each for VS & PS
 const VERTEX_BUFFER_MAX_BYTES: u64 = 16 * 1024 * 1024;
 
 pub struct XenosPipeline {
+    /// Interpreter pipeline with the legacy fixed (alpha-blend) state. Kept as
+    /// the default; per-state variants are built lazily in `interp_cache`.
     pipeline: wgpu::RenderPipeline,
+    /// iterate-3Y: the interpreter WGSL module, retained so per-render-state
+    /// interpreter pipelines can be compiled on demand.
+    interp_shader: wgpu::ShaderModule,
+    /// iterate-3Y: interpreter pipelines keyed on the per-draw `RenderState`
+    /// (blend + write mask), so flat/alpha/opaque draws composite correctly
+    /// even when their (vs,ps) didn't translate.
+    interp_cache: std::collections::HashMap<RenderState, wgpu::RenderPipeline>,
     draw_ctx_buffer: wgpu::Buffer,
     constants_buffer: wgpu::Buffer,
     vs_ucode_buffer: wgpu::Buffer,
@@ -78,7 +229,12 @@ pub struct XenosPipeline {
     /// so every (vs, ps) pair gets compiled once and re-used for every
     /// subsequent draw. Interpreter pipeline remains the fallback.
     pipeline_layout: wgpu::PipelineLayout,
-    translated_cache: std::collections::HashMap<(u32, u32), wgpu::RenderPipeline>,
+    /// iterate-3Y: cached translator pipelines keyed on the shader pair AND the
+    /// per-draw render state, so the same (vs,ps) with different blend/mask
+    /// composites correctly. The translated WGSL module is itself cached per
+    /// (vs,ps) so re-translation only happens once.
+    translated_cache: std::collections::HashMap<(u32, u32, RenderState), wgpu::RenderPipeline>,
+    translated_modules: std::collections::HashMap<(u32, u32), wgpu::ShaderModule>,
     pub target_format: wgpu::TextureFormat,
 }
 
@@ -193,7 +349,9 @@ impl XenosPipeline {
             draw_index: 0,
             vertex_count: 3,
             prim_kind: 4,
-            _pad: 0,
+            vertex_base_dwords: 0,
+            ndc_scale: [0.0, 0.0],
+            ndc_offset: [0.0, 0.0],
         };
         let draw_ctx_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
             label: Some("xenos draw ctx"),
@@ -242,8 +400,13 @@ impl XenosPipeline {
             usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
             view_formats: &[],
         });
-        // Magenta (255, 0, 255, 255) so a missing-texture read visibly stands
-        // out on-screen when the interpreter does sample it.
+        // iterate-3Y: transparent black (0,0,0,0). When a textured draw's
+        // real texture can't be resolved (e.g. its sampler slot is shadowed by
+        // a vertex-fetch constant), sampling a *transparent* texel makes the
+        // draw a no-op under its real premultiplied-alpha blend — instead of
+        // fabricating an opaque magenta that overpaints everything (the old
+        // debug stub). This removes a fake rather than adding one: we never
+        // invent visible pixels for an unresolved texture.
         queue.write_texture(
             wgpu::ImageCopyTexture {
                 texture: &dummy_tex,
@@ -251,7 +414,7 @@ impl XenosPipeline {
                 origin: wgpu::Origin3d::ZERO,
                 aspect: wgpu::TextureAspect::All,
             },
-            &[0xFFu8, 0x00, 0xFF, 0xFF],
+            &[0x00u8, 0x00, 0x00, 0x00],
             wgpu::ImageDataLayout {
                 offset: 0,
                 bytes_per_row: Some(4),
@@ -359,6 +522,8 @@ impl XenosPipeline {
 
         Self {
             pipeline,
+            interp_shader: shader,
+            interp_cache: std::collections::HashMap::new(),
             draw_ctx_buffer,
             constants_buffer,
             vs_ucode_buffer,
@@ -371,31 +536,22 @@ impl XenosPipeline {
             dummy_view,
             pipeline_layout: layout,
             translated_cache: std::collections::HashMap::new(),
+            translated_modules: std::collections::HashMap::new(),
             target_format,
         }
     }
 
-    /// P7 — does the translator cache already have a pipeline for this
-    /// (vs, ps) pair?
+    /// P7 — has the translator already produced a WGSL *module* for this
+    /// (vs, ps) pair? (A per-render-state pipeline may still need building.)
     pub fn has_translated(&self, vs_blob_key: u32, ps_blob_key: u32) -> bool {
-        self.translated_cache
+        self.translated_modules
             .contains_key(&(vs_blob_key, ps_blob_key))
     }
 
-    /// P7 — fetch a cached translator pipeline. `None` if not yet built.
-    pub fn translated_pipeline(
-        &self,
-        vs_blob_key: u32,
-        ps_blob_key: u32,
-    ) -> Option<&wgpu::RenderPipeline> {
-        self.translated_cache
-            .get(&(vs_blob_key, ps_blob_key))
-    }
-
-    /// P7 — compile a translator-produced WGSL module into a
-    /// `wgpu::RenderPipeline` and insert it into the cache keyed on
-    /// `(vs_blob_key, ps_blob_key)`. Returns `true` on success. Duplicate
-    /// inserts are no-ops. Emits `gpu.shader.compile_ok` on success.
+    /// P7 — compile a translator-produced WGSL module and cache it keyed on
+    /// `(vs_blob_key, ps_blob_key)`. The actual `RenderPipeline` (which also
+    /// depends on the per-draw blend/mask state) is built lazily by
+    /// [`render_one_translated`]. Returns `true` on success.
     pub fn insert_translated(
         &mut self,
         device: &wgpu::Device,
@@ -404,7 +560,7 @@ impl XenosPipeline {
         wgsl: &str,
     ) -> bool {
         let key = (vs_blob_key, ps_blob_key);
-        if self.translated_cache.contains_key(&key) {
+        if self.translated_modules.contains_key(&key) {
             return true;
         }
         let shader = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
@@ -420,31 +576,42 @@ impl XenosPipeline {
                 return false;
             }
         };
+        self.translated_modules.insert(key, shader);
+        metrics::counter!("gpu.shader.compile_ok").increment(1);
+        true
+    }
+
+    /// iterate-3Y: ensure a translator pipeline exists for `(vs,ps,rstate)`,
+    /// building it from the cached module + the per-draw color/blend target.
+    fn ensure_translated_for_state(
+        &mut self,
+        device: &wgpu::Device,
+        vs_key: u32,
+        ps_key: u32,
+        rstate: RenderState,
+    ) -> bool {
+        let pkey = (vs_key, ps_key, rstate);
+        if self.translated_cache.contains_key(&pkey) {
+            return true;
+        }
+        let Some(module) = self.translated_modules.get(&(vs_key, ps_key)) else {
+            return false;
+        };
+        let target = rstate.color_target(self.target_format);
         let pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
             label: Some("xenos translated pipeline"),
             layout: Some(&self.pipeline_layout),
             vertex: wgpu::VertexState {
-                module: &shader,
+                module,
                 entry_point: "vs_main",
                 compilation_options: Default::default(),
                 buffers: &[],
             },
             fragment: Some(wgpu::FragmentState {
-                module: &shader,
+                module,
                 entry_point: "fs_main",
                 compilation_options: Default::default(),
-                targets: &[Some(wgpu::ColorTargetState {
-                    format: self.target_format,
-                    blend: Some(wgpu::BlendState {
-                        color: wgpu::BlendComponent {
-                            src_factor: wgpu::BlendFactor::SrcAlpha,
-                            dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
-                            operation: wgpu::BlendOperation::Add,
-                        },
-                        alpha: wgpu::BlendComponent::OVER,
-                    }),
-                    write_mask: wgpu::ColorWrites::ALL,
-                })],
+                targets: &[Some(target)],
             }),
             primitive: wgpu::PrimitiveState {
                 topology: wgpu::PrimitiveTopology::TriangleList,
@@ -460,30 +627,78 @@ impl XenosPipeline {
             multiview: None,
             cache: None,
         });
-        self.translated_cache.insert(key, pipeline);
-        metrics::counter!("gpu.shader.compile_ok").increment(1);
+        self.translated_cache.insert(pkey, pipeline);
         true
     }
 
-    /// Render one draw with the translator-produced pipeline instead of
-    /// the interpreter. Mirrors [`render_one`] except the bound pipeline
-    /// is swapped for `pipeline`.
-    pub fn render_one_with_pipeline(
-        &self,
+    /// iterate-3Y: ensure an interpreter pipeline exists for `rstate`.
+    fn ensure_interp_for_state(&mut self, device: &wgpu::Device, rstate: RenderState) {
+        if self.interp_cache.contains_key(&rstate) {
+            return;
+        }
+        let target = rstate.color_target(self.target_format);
+        let pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+            label: Some("xenos interp pipeline (per-state)"),
+            layout: Some(&self.pipeline_layout),
+            vertex: wgpu::VertexState {
+                module: &self.interp_shader,
+                entry_point: "vs_main",
+                compilation_options: Default::default(),
+                buffers: &[],
+            },
+            fragment: Some(wgpu::FragmentState {
+                module: &self.interp_shader,
+                entry_point: "fs_main",
+                compilation_options: Default::default(),
+                targets: &[Some(target)],
+            }),
+            primitive: wgpu::PrimitiveState {
+                topology: wgpu::PrimitiveTopology::TriangleList,
+                strip_index_format: None,
+                front_face: wgpu::FrontFace::Ccw,
+                cull_mode: None,
+                polygon_mode: wgpu::PolygonMode::Fill,
+                unclipped_depth: false,
+                conservative: false,
+            },
+            depth_stencil: None,
+            multisample: wgpu::MultisampleState::default(),
+            multiview: None,
+            cache: None,
+        });
+        self.interp_cache.insert(rstate, pipeline);
+    }
+
+    /// iterate-3Y: render one draw through the translator pipeline built for
+    /// this draw's render state. Returns `false` if no module is cached for
+    /// `(vs,ps)` (caller should fall back to the interpreter).
+    pub fn render_one_translated(
+        &mut self,
+        device: &wgpu::Device,
         queue: &wgpu::Queue,
         encoder: &mut wgpu::CommandEncoder,
         target_view: &wgpu::TextureView,
         req: DrawRequest,
-        pipeline: &wgpu::RenderPipeline,
-    ) {
+        vs_key: u32,
+        ps_key: u32,
+        rstate: RenderState,
+    ) -> bool {
+        if !self.ensure_translated_for_state(device, vs_key, ps_key, rstate) {
+            return false;
+        }
         let cb = DrawConstants {
             draw_index: req.draw_index,
             vertex_count: req.vertex_count.max(3),
             prim_kind: req.prim_kind,
-            _pad: 0,
+            vertex_base_dwords: req.vertex_base_dwords,
+            ndc_scale: req.ndc_scale,
+            ndc_offset: req.ndc_offset,
         };
         queue.write_buffer(&self.draw_ctx_buffer, 0, bytemuck::bytes_of(&cb));
-
+        let pipeline = self
+            .translated_cache
+            .get(&(vs_key, ps_key, rstate))
+            .expect("just ensured");
         let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
             label: Some("xenos translated draw"),
             color_attachments: &[Some(wgpu::RenderPassColorAttachment {
@@ -503,6 +718,7 @@ impl XenosPipeline {
         pass.set_bind_group(1, &self.tex_bind_group, &[]);
         let rounded = req.vertex_count.div_ceil(3) * 3;
         pass.draw(0..rounded.max(3), 0..1);
+        true
     }
 
     /// Number of distinct translator pipelines cached. Surfaced to the HUD.
@@ -594,22 +810,34 @@ impl XenosPipeline {
         queue.write_buffer(&self.vertex_buffer, 0, &bytes[..bytes.len().min(max)]);
     }
 
-    /// Render one captured draw.
+    /// Render one captured draw through the interpreter, using the per-draw
+    /// `rstate` (blend/write-mask) so flat draws composite correctly even
+    /// when their (vs,ps) didn't translate. `RenderState::OPAQUE` reproduces
+    /// the legacy fixed behaviour for procedural/synthetic draws.
     pub fn render_one(
-        &self,
+        &mut self,
+        device: &wgpu::Device,
         queue: &wgpu::Queue,
         encoder: &mut wgpu::CommandEncoder,
         target_view: &wgpu::TextureView,
         req: DrawRequest,
+        rstate: RenderState,
     ) {
+        self.ensure_interp_for_state(device, rstate);
         let cb = DrawConstants {
             draw_index: req.draw_index,
             vertex_count: req.vertex_count.max(3),
             prim_kind: req.prim_kind,
-            _pad: 0,
+            vertex_base_dwords: req.vertex_base_dwords,
+            ndc_scale: req.ndc_scale,
+            ndc_offset: req.ndc_offset,
         };
         queue.write_buffer(&self.draw_ctx_buffer, 0, bytemuck::bytes_of(&cb));
 
+        let pipeline = self
+            .interp_cache
+            .get(&rstate)
+            .expect("just ensured");
         let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
             label: Some("xenos draw"),
             color_attachments: &[Some(wgpu::RenderPassColorAttachment {
@@ -624,7 +852,7 @@ impl XenosPipeline {
             timestamp_writes: None,
             occlusion_query_set: None,
         });
-        pass.set_pipeline(&self.pipeline);
+        pass.set_pipeline(pipeline);
         pass.set_bind_group(0, &self.bind_group, &[]);
         pass.set_bind_group(1, &self.tex_bind_group, &[]);
         let rounded = req.vertex_count.div_ceil(3) * 3;
@@ -638,6 +866,6 @@ mod tests {
 
     #[test]
     fn draw_constants_layout_matches_wgsl_uniform() {
-        assert_eq!(std::mem::size_of::<DrawConstants>(), 16);
+        assert_eq!(std::mem::size_of::<DrawConstants>(), 32);
     }
 }