79d0026a31
Windowed (--ui) mode forces inline GPU, so VdSwap's per-swap UI publish runs on the emulation thread — profiled at ~87% of it (~4 MIPS effective vs ~35 headless). This lands the low-risk, headless-untouched wins (golden n200m byte-identical): - A.1 (biggest): VdSwap frontbuffer detile now uses one bounded GuestMemory ::read_bulk instead of ~3.7MB of per-byte read_u8 through the MMIO handler (~15 ms/swap). Bounds-checked to stay in the committed backing window. - A.4: XENIA_PRESENT_MODE (immediate|mailbox|fifo) + XENIA_FRAME_LATENCY knobs (render.rs); default (Mailbox-else-Fifo, latency 2) unchanged. - A.3a: cache parse_shader/pack_for_wgsl per blob key on RenderState instead of re-parsing every draw every frame (blobs are immutable) — the movie-relevant UI-thread win. - A.2: publish shader-blob map to the UI only when it changed (shader_blobs_version on GpuSystem; publish_xenos_assets blobs arg is now Option, None = keep previous). Constants still published every swap. Deferred (profiling-justified — target the measured bottlenecks, not these): - A.2 texture-gate, A.3b bind-group cache: zero benefit for the movie (its texture keys rotate every frame → always-miss) + staleness/leak risk. - A.3c submit-batching: the UI thread's bottleneck is the vsync-blocked present, not per-draw submits; GPUBUG-111 regression risk not justified. - A.5 (threaded GPU under --ui): the structural win; separate follow-up (needs the publish bridge moved to the worker + human visual verification). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>