xenia-rs/audit-runs/phase-host-audio-eager/re-validation.md

Phase Host-Audio-Eager — Re-validation (2026-05-19)

Progression metric (primary gate)

metric	pre-fix baseline	post-fix	delta
swaps	1	1	0
draws	0	0	0

The progression metric did NOT move. Despite landing the eager-seed implementation cleanly with 3× reproducibility, neither swaps nor draws advanced. This matches the prior agent's diagnosis: the audio worker ordering issue is real, but the deeper root cause is voice-struct state divergence — the audio callback at 0x824D6640 in ours blocks on KeWaitForMultipleObjects([0x82928B04, 0x82928AE0]) immediately because the voice struct at [r31+356] reads 0x01 (ours) vs 0x00 (canary). Pre-seeding 8 fires lets try_inject_audio_callback deliver the first callback earlier, but the callback still blocks on the same guest dispatchers — fires 2-8 sit in the queue because interrupts.is_in_callback() stays true.

3× determinism

73e99d60029128b4d5c3dd98e540457d82a52b8a962e7495132be2be31411aca  /tmp/digest_eager_1.json
73e99d60029128b4d5c3dd98e540457d82a52b8a962e7495132be2be31411aca  /tmp/digest_eager_2.json
73e99d60029128b4d5c3dd98e540457d82a52b8a962e7495132be2be31411aca  /tmp/digest_eager_3.json

All three cold runs produce byte-identical digest JSON. The seed-at-register implementation is fully deterministic in lockstep mode (the ticker accumulator gets pre-populated synchronously inside the register handler, no host-thread non-determinism).

Digest JSON

{
  "instructions": 50000007,
  "imports": 40390,
  "unimpl": 0,
  "draws": 0,
  "swaps": 1,
  "unique_render_targets": 0,
  "shader_blobs_live": 0,
  "texture_cache_entries": 0
}

imports/unimpl unchanged from the C+22 baseline (40390/0).

Phase B invariant

image_loaded_sha256 = ea8d160e9369328a5b922258a92113efb8d7ce3e1a5c12cc521e375985c91c18

UNCHANGED — Phase B is not affected by audio-runtime changes.

Per-chain matched-prefix

100M-instruction cold trace vs canary baseline (xenia-rs/audit-runs/phase-d-stage1/canary-cvaroff-trunc.jsonl — pre Phase D D-extension absorber, so main reads at the C+18 102,424 value, NOT the post-D-extension 105,046).

chain	pre-fix	post-fix	delta	first divergence
canary tid=4 → ours tid=11	11	11	0	(preserved)
canary tid=6 → ours tid=1	102,424	102,424	0	NtQueryFullAttributesFile (C+18-era)
canary tid=7 → ours tid=2	32	32	0	(preserved)
canary tid=12 → ours tid=7	4	4	0	C+23 idx=4
canary tid=14 → ours tid=9	41	41	0	(no advance — primary target)
canary tid=15 → ours tid=10	16	16	0	(no advance — primary target)

The two primary targets (tid=14→9 and tid=15→10) were the audio worker guest threads spinning on the uninitialized voice struct. Their matched-prefix did NOT advance.

Kernel tests

• Pre: 217 passed
• Post: 221 passed (+4 new seed_fires_for_* tests)
• Failures: 0
• All existing tests pass

Build

cargo build --release clean. One pre-existing dead-code warning unrelated to this fix.

Total LOC

file	added	removed
crates/xenia-kernel/src/xaudio.rs	86	0
crates/xenia-kernel/src/exports.rs	18	5
total	104	5

Net ~100 LOC, of which ~60 LOC are tests + doc comments. Engine logic delta is ~25 LOC.

Conclusion

The implementation lands cleanly:

• 3× cold-deterministic
• Phase B unchanged
• All tests pass
• Sister chains preserved

But the progression metric (swaps/draws) did NOT move. This is an HONEST NEGATIVE RESULT: the eager-seed approach addresses the symptom (ticker delays the first callback) but not the root cause (the callback at 0x824D6640 still blocks on guest dispatchers that only tid=9/10 can signal, and tid=9/10 are stuck on a voice-struct field that the callback would need to clear — but doesn't, because canary's callback takes a DIFFERENT control-flow path that doesn't reach the KeWaitForMultipleObjects in the first place).

The deeper fix requires either:

• Identifying the guest write that initializes [r31+356] to 0 in canary's boot path and ensuring ours produces the same write.
• A true host-side audio worker thread that can run the callback in a host context (substantial threading-model rework).

Both are out of scope for this session per the brief's "Don't widen scope" tripstone.