Skip to content

emulation-bug-hunt-todos.md

Latest commit

 

History

History
490 lines (456 loc) · 26.8 KB

emulation-bug-hunt-todos.md

File metadata and controls

490 lines (456 loc) · 26.8 KB

Emulation Bug Hunt TODOs

This tracks actionable issues found while auditing old inline comments that describe guesses, timing gaps, or known emulation shortcuts.

Cross-cutting finding from cluster investigation (2026-04-30)

A round of parallel sub-agent investigation snapshotted DSP RAM, 68K regs, TOM/JERRY/DAC regs, low main RAM, and rendered framebuffers at multiple frames for the still-broken cart titles in both real-BIOS and HLE modes. Headline result: 4 of 5 hang/crash titles (Hyper Force, Iron Soldier, Ruiner Pinball, Super Burnout) have identical or near-identical behavior in BIOS vs HLE — same stuck PC, same nonblack pixel count, same DSP state. So those titles are real emulation bugs, not HLE-init issues. Real BIOS doesn't fix them either.

The narrowest reproducible per-title clue from the snapshots:

  • Wolfenstein 3D (audio, both BIOS and HLE):
    • HLE: completely silent (RMS=0, first-audio=-1) for the entire run (verified up to frame 1800 / 30s).
    • BIOS: brief BIOS-chime audio at frames 34-~600, then silent forever. The BIOS-chime is not Wolf3D's own audio, it's the BIOS's startup tone played before the cart takes over. Wolf3D's own game audio never starts in either mode.
    • User-reported (real RetroArch on iOS): also no audio in either mode, plus a BIOS-specific quirk that pressing A/B during the BIOS logo skips initialization and Wolf3D then fails to boot entirely (no other game does that).
    • Root cause: Wolf3D's DSP escapes work RAM by frame 1800. At that point dsp_pc=0x000003FA (HLE) / 0x00181C43 (BIOS) — both in main RAM, not DSP work RAM (0xF1B000-0xF1CFFF). Atari Karts HLE keeps DSP at 0xF1B3FA correctly with active LTXD/RTXD. So Wolf3D's DSP code (whether BIOS-loaded or cart-loaded) has a JUMP-through-register instruction where the register holds a garbage value, sending the DSP into main-RAM/68K code. Tried memcpy'ing the BIOS DSP audio engine (jaguarBootROM[0x214E..0x2916], 1992 bytes) into DSP RAM and starting the DSP with D_PC = engine entry / mainloop / DSPGO=1 — same DSP-escape symptom because the engine reads DSP registers we never initialize and uses them as jump targets. Reverted the engine copy; needs full DSP register-bank state replication.
  • Skyhammer (HLE, audio): agent's snapshot showed 68K at 0x008022EE for frames 1-60. Followup boot_timeline at frames 60/300/600/1200/3600/7200 shows PC actually progresses through cart code (0x832F4A → 0x8B6378 → 0x1644C → 0x16610 → 0x8AC7E2 → 0x8AC930) — Skyhammer's 68K is NOT stuck. The DBF loop at 0x22EE is just a long delay (D0=0xFFFFFF, ~167M cycles, ~6 seconds at 13.3 MHz). Audio clipping is purely a DSP-side / I2S issue — the cart's main 68K thread runs fine. Cart-disassembly at the DBF address is not the right approach; root cause is in JERRY/DSP timing or HLE audio-engine missing state.
  • Raiden (HLE, won't boot): cart's startup at 0x802000 copies game code from cart 0x802026..0x802986 (0x960 bytes) into main RAM at 0x180000+, then JMPs there. Code runs initial setup (LEA into A4 = 0x1803B2) and then enters a polling loop at 0x18014A: TST.B $2C7(A4); BEQ.S -4 — i.e. spin forever until RAM[0x180679] becomes non-zero. Some interrupt handler is supposed to set that byte; our HLE generic-RTE stub at 0x404 just RTE's so the byte never gets set and the cart hangs. 68K stack pointer A7 stays stuck at 0x001FFFFC (one push deep from initial 0x00200000), confirming no interrupts are actually firing during the spin (so it's not "exception double-fault" — agent's earlier hypothesis was based on wrong dlsym dereference and is incorrect). Fix path: cart probably installs its own IRQ handlers via the BSRs at 0x180000 and 0x180004 BEFORE the poll loop; need to trace what those BSRs do and identify why interrupts aren't arriving (TOM video IRQ enable, JERRY IRQ enable, vector base).
  • Ruiner Pinball: identical 0x809CAE-stuck PC and 0% nonblack pixels in BIOS and HLE. Cart never gets past initialization. Disassembling around the stuck address shows a routine at 0x9CA0 that does CMPI.L #0, $00005B18; BEQ +6; JMP $00802000, plus a separate routine at 0x2248 that calls a function pointer at $0000402C, runs JSR $00802380, then writes MOVE.L #1, $00005B18 only if RAM[$4068] has bits 0+16 set and bit 9 clear. Probing those addresses each frame for 600 frames in both BIOS and HLE shows: $402C and $4068 stay 0 forever in both modes, while $5B18 cycles through various non-zero values (PRNG / cart-side scratch). So the cart is missing the initialization that should populate $402C (function pointer) and $4068 (interrupt-flags accumulator). Real BIOS does not provide these — implies the cart itself is responsible, but a precondition for that init is failing (likely an interrupt that should fire early in boot but doesn't). v2.3.0 work: trace the cart's interrupt-handler installation path and identify what precondition needs to hold for the init routine to run.
  • Hyper Force, Iron Soldier, Super Burnout: matching pixel counts and PC ranges in BIOS/HLE. Engine-level bugs, not init.
  • NBA Jam TE / Towers II / Tempest 2000: empirical flicker score 2.6× / 2.6× / 1.3× the Atari Karts baseline (per-pixel temporal stddev across a 16-frame window). Confirms the symptom and gives a regression-watcher metric.

A more accurate JaguarReset HLE init now writes SCLK=0x13 and SMODE=0x15 (matching what the BIOS audio engine ends up programming — INTERNAL + WSEN + FALLING; previous defaults 0x08 / 0x01 were the BIOS pre-engine values). Verified Atari Karts negative control still clean; did NOT fix Skyhammer / IS2 audio clipping — the DBF-loop diagnosis above explains why a single register-init tweak isn't enough.

Raw investigation data left in /tmp/: /tmp/cluster_findings.md (summary), /tmp/hle_diff_*.md, /tmp/flick_*.txt+.ppm, /tmp/boot_logs/*.txt, /tmp/{hyper,iron,hover,ruiner,super}_*_1800.png.

Recently Addressed

  • TOM frame rollover now follows the VP register instead of hard-coded NTSC/PAL field lengths. This targets visible "screen slap" jumps when games program custom video timing.
  • TOM visible-line rendering no longer computes framebuffer pointers for offscreen halflines.
  • Event scheduling now treats negative residual times as "due now" instead of feeding negative microseconds into CPU/GPU/DSP cycle conversion.
  • JERRY JINTCTRL word writes are now decoded only at $F10020, so writes to the adjacent $F10022 word cannot alter the 68K interrupt mask or pending latches.
  • Libretro geometry changes now apply after the frame rendered with the previous pitch is submitted, avoiding a one-frame pitch mismatch when games reprogram TOM display registers.
  • The dead objectp_running flag and its misleading VMODE write side effect have been removed; OP execution is driven directly by halfline display state.
  • The unused pipelined DSPHandleIRQs export has been removed; live DSP interrupt dispatch goes through the non-pipelined handler used by DSPSetIRQLine and the IMASK-cleared path.
  • OP scaled bitmap clipping now keeps fractional phrase width for all edge cases, avoiding divide-by-zero when small hscale values truncate an integer-scaled phrase width to zero.
  • OP scaled bitmap rendering now starts the horizontal scale phase at zero for 1:1 hscale=$20 output and applies firstPix to the first source phrase, treats iwidth == 0 as one phrase, keeps the visible edge pixel for reflected left-edge objects, and handles magnified hscale source stepping, with direct 4 BPP coverage for 1:1, 3:2, 2:1, left/right-edge clipping, and reflected edge clipping in test_hle_bios.
  • OP fixed bitmap rendering now honors firstPix for 2/4/16/24 BPP paths, not just 1/8 BPP, and avoids applying it again after clipping skips whole source phrases, with direct 4 BPP coverage in test_hle_bios.
  • TOM interrupt sources now latch pending status even when their CPU enable bits are clear; enabling a pending source asserts the 68K IPL2 line, with direct video IRQ latch coverage in test_hle_bios.
  • Headerless raw homebrew loading is now conservative but supported for recognizable startup patterns at inferred $4000, $20000, or $802000 bases; unknown raw files still fail instead of booting invalid RAM.
  • JERRY SSI SSTAT reads now report the modeled status bits instead of the generic unmapped $FFFF placeholder.
  • make test now includes test_dsp_unit, covering DSP interrupt dispatch, priority, return-address stack push, register banking, and execution basics.
  • test_hle_bios now covers HLE workspace state, exception vectors, I2S defaults and SSTAT readback, JERRY JINTCTRL decode, OP scaled clipping, deferred geometry updates, PAL timing, and custom VP rollover.
  • make test now includes event queue coverage for zero/negative-time event handling.
  • make test now runs test_audio_clipping, an automated boot-time audio clipping detector. The test captures the libretro audio batch output, measures saturation density, longest run at +/-32767, and sustained RMS, and asserts on a negative control (Atari Karts) plus two known-broken regression watchers (Skyhammer, Iron Soldier 2) flagged with --expect-clipping so the test goes red the day a fix lands and the bug disappears. A/B-tested against libretro/master: Skyhammer's clipping is pre-existing (master clips harder, ~34% sample-saturation density vs ~25% on this branch) so it is not a regression introduced here.

Game compatibility (v2.2.0)

Status of titles called out in the libretro tracker (issue #38 sub-list). Verified against private ROMs unless otherwise noted.

Fixed in v2.2.0

  • Air Cars — title screen is no longer cut off.
  • Atari Karts — vertical bar artifact on the side during gameplay is gone.
  • Battlesphere Gold — menu black-screen on real-BIOS path is fixed (a separate fast-blitter menu rendering glitch still exists; see below).
  • Club Drive — title screen is no longer cut off, missing textures restored.
  • Cybermorph — missing/glitched textures fixed; the digitised Skylar voice also plays correctly now.
  • Doom — vertical bar gone; the legacy "Doom resolution hack" core option is removed (proper PWIDTH pixel replication makes it unnecessary).
  • Flashback - Quest for Identity — full-screen display restored (was anchored to the left).
  • I-War — title cut-off and gameplay glitches fixed on the accurate blitter (fast blitter has a separate flickering issue; see below).
  • Kasumi Ninja — title-screen flicker, missing textures, and post-title freezes fixed (some stages still show vertical tearing 3/4 down the screen, regardless of BIOS — see below).
  • Missile Command 3D — minor flickering / glitches resolved.
  • Pinball Fantasies — intro screens no longer cut off.
  • Powerdrive Rally — intro screens no longer cut off; new-game crash gone.
  • Skyhammer — gameplay no longer freezes (audio is clipped/loud, see below).
  • Supercross 3D — full-screen display restored, textures readable.
  • Syndicate — pink-patches-in-text fixed on the accurate blitter (fast blitter still has the original issue; see below).
  • Trevor McFur in the Crescent Galaxy — loading-screen cut-off fixed.
  • Val D'Isere Skiing & Snowboarding — full-screen display restored (a player-vs-snow z-order regression introduced; see below).
  • Zero 5 — flickering resolved.
  • Zoop — game no longer crashes on launch.
  • HLE BIOS bridge — Raiden / Kasumi Ninja and similar titles that previously needed a real BIOS image now boot under HLE.

Still broken / regressed (track these into v2.3.0)

  • Battle Sphere (accurate blitter): the targeting reticle drawn over enemies leaves dark/dotted artefacts where the framebuffer should show through. Fast blitter renders this correctly. Likely involves phrase-mode byte_merge against DCOMPEN / BCOMPEN comparators with BKGWREN enabled — the inhibited bytes are taking the DSTDATA register value instead of the framebuffer pixel. A first attempt to widen phrase_mode && !dsten to also read framebuffer when comparators are on did not fix the bug, so the case is more nuanced than the obvious gate. Needs a logged COMMAND/A1/A2 register dump from one bad blit.
  • Battlesphere Gold (fast blitter): menu rendering glitches (separate from the accurate-blitter reticle issue) regardless of BIOS. Fast-blitter-only.
  • Fight for Life: flickering graphics, mainly in the menu — looks like black bars scrolling over text/button texture layers. Some layer compositing ordering is off.
  • Hover Strike: title and graphics fixed, but the game crashes during gameplay, with or without real BIOS.
  • Hyper Force: black screen after title; doesn't progress with or without real BIOS.
  • Iron Soldier: boots through title and character select, then black-screens (music keeps looping, looks like a video-mode switch that fails). With or without real BIOS.
  • Iron Soldier 2: title now displays correctly, but the music is clipped/over-loud on the title and the screen black-screens after character selection.
  • NBA Jam Tournament Edition: occasional screen jumping. Same pattern as White Men Can't Jump (both High Voltage Software); likely an engine-shared timing/IRQ-phase issue. Tracked separately.
  • Raiden: still requires real BIOS — won't boot on HLE.
  • Ruiner Pinball: won't boot regardless of BIOS, PAL/NTSC, or blitter mode.
  • Super Burnout: new-game now starts but doesn't render correctly — only the background layer draws (no cars, scenery, or HUD). Audio is present.
  • Tempest 2000: rare flickering still occurs; the accurate blitter is also significantly slower than fast on this title, enough to drop below full speed on hardware where Jaguar should comfortably run. Performance-side TODO.
  • Towers II: unplayable due to flickering/flashing. PAL mode is worse, and PAL audio is broken — looks like a timing issue.
  • White Men Can't Jump: boots, but lots of flickering; can't get past main menu on HLE; real BIOS doesn't boot the splash at all. (See the more-detailed entry in High Priority below.)
  • Wolfenstein 3D: if the real BIOS is allowed to fully boot before pressing a button the game starts (audio still broken); exiting BIOS early black-screens. HLE is broken.
  • Val D'Isere (regression): the player sprite z-order with the snow layer is reversed — snow draws over the skier unless the player jumps above the horizon line. New regression in v2.2.0.
  • I-War (fast blitter): title cut-off returns when the fast blitter is enabled. Accurate blitter has a separate floor-layer flicker during gameplay.
  • Kasumi Ninja (some stages): vertical tearing 3/4 down the screen on certain stages, regardless of BIOS.
  • Syndicate (fast blitter): pink-patches-in-text remains on the fast blitter (accurate blitter is fine).
  • Skyhammer / Iron Soldier 2 audio clipping: boot-time audio is a saturated square wave (sustained +/-32767, ~25% saturation density on Skyhammer, ~21% on IS2). Detected automatically by test_audio_clipping.
    • Repro: load with HLE BIOS, run ~300 frames, alternates +/-32767 with rare in-between values.
    • Critical clue (2026-04-29): with real BIOS, Skyhammer audio is clean (RMS ~3987, 0% saturation). With HLE, it's the saturated square wave. So the bug is HLE-init delta, not DSP arithmetic. Atari Karts works on HLE — its DSP code is self-contained, while Skyhammer / IS2 likely depend on BIOS-loaded DSP audio engine state. (Atari Karts was A/B-tested as the negative control.)
    • Pre-existing: A/B-tested against libretro/master — master clips harder (~34% on Skyhammer) so this is a long-standing bug, not a regression introduced by this PR.
    • DSP-state diff (2026-04-29): snapshotted DSP RAM at frame 5/10 in BIOS vs HLE mode and binary-searched the embedded jaguarBootROM blob for the matching prefix:
      • The BIOS pre-loads a 1992-byte DSP audio engine at jaguarBootROM[0x214E .. 0x2916] into DSP RAM offset 0 and starts the DSP (DSPGO=1).
      • Engine prefix at offset 0: 98 00 B0 30 00 F1 D0 00 E4 00 E4 00 E4 00 E4 00 (MOVEI #$F1D000, R0 — wavetable ROM ptr — then NOP slots).
      • But copying this engine into DSP RAM and starting the DSP in HLE init does NOT fix the clipping. Both Skyhammer and IS2 OVERWRITE the engine with their own DSP code by ~frame 30, so having it pre-loaded is moot for them.
      • Atari Karts is unaffected by the engine copy (verified — same RMS 410.8, 0% saturation).
      • The DSP code Skyhammer loads in HLE-mode at frame 175 is dramatically different from the code it loads in BIOS-mode at frame 175 (DSP RAM contents diverge across ~95% of the audio engine area). So Skyhammer's 68K code is reading something at boot to choose which DSP audio routine to load, and HLE provides a different value than BIOS.
    • Next steps: trace what Skyhammer's 68K code reads early-boot from BIOS-area (0xE00000+), low main RAM (0x0000-0x0800), or BIOS-installed exception vectors. The BIOS leaves a vector table with handler addresses like 06066xxx, 06067xxx; HLE installs RTE stubs at different addresses. If Skyhammer JSRs through one of these vectors to a BIOS audio-init routine, our RTE stub returns immediately and the routine never runs. That's the most plausible remaining hypothesis.
    • Diagnostic tools: /tmp/dsp_diff.c and /tmp/dsp_snapshot.c in the conversation log capture DSP RAM and key DAC/DSP registers via dlsym — re-buildable from the recipe in docs/test-infrastructure.md if needed.

Performance

  • Accurate blitter is significantly slower than fast on several titles (Tempest 2000 is the loudest). The fast-blitter and SIMD paths cover the common case; the scalar accurate path needs more inner-loop tightening or an SIMD-accelerated DCOMPEN / BCOMPEN variant.

High Priority

  • src/tom/op.c: replace the Object Processor runaway-list guard with a real resumable scheduler. The current OP_RUNAWAY_GUARD_OBJECTS limit prevents malformed lists from hanging the emulator, but it does not model OP cycle consumption or overloaded-list suspend/reentry timing.
  • White Men Can't Jump: private-ROM tracing shows the visible jumps happen with a stable OLP and object-list contents. The jumping object is a fixed bitmap consumed by OP write-back and restored outside direct 68K stores, through the 68K VBI path. Bad frames occur when the video interrupt lands while the 68K SR masks level-2 IRQs, delaying the object restore until active display (VC=$0866 in the captured repro). Suppressing OP write-back hides the jump but contradicts hardware behavior. TOM timer IRQs are the immediate collision source: disabling TOM PIT interrupts removes the boot-logo jump, while alternate PIT reload formulas and line-buffer clearing probes made the frame sequence worse. A real TOM fix landed for INT1 byte reads/writes exposing and clearing pending IRQ sources consistently with word accesses, but WMCJ still needs 68K/TOM interrupt phase accuracy rather than bitmap geometry.
  • src/tom/op.c: continue auditing scaled bitmap semantics beyond the small-hscale, firstPix, 1:1 phase, iwidth == 0, and reflected edge fixes. Additional non-integer ratios and larger reflected phrase alignment cases still need repro or hardware coverage because they affect road/ground rendering.

Medium Priority

  • src/tom/tom.c: replace hard-coded visible-window constants with values derived from TOM timing registers where possible. This should be tested against NTSC, PAL, and games that reprogram HDB/HDE/VDB/VDE.
  • src/tom/op.c: document or test the transparent index / encoded black behavior in 4/8 BPP paths. Comments disagree about whether index 0 is transparent in all relevant modes.
  • src/jerry/dac.c: continue SSI/I2S modeling beyond basic SSTAT readback, including slave-clock timing and CD audio word-strobe behavior.
  • src/jerry/dsp.c: use the new IRQ return-address coverage before changing live non-pipelined dispatch semantics.

Lower Priority / CD Branch

  • src/cd/cdintf.c: many CD interface methods remain stubs.
  • src/cd/cdrom.c: FIFO, I2S-to-JERRY, DSA command timing, and interrupt generation have several comments marked as approximate or hard-coded.
  • src/core/filedb.c: cartridge/file-type detection still has known edge cases noted by inline comments.

v2.3.0 follow-up notes (cleanup + investigations)

These came out of the v2.2.0 manual pre-merge review. None block shipping v2.2.0; capture them so they don't get lost.

Removed code we should remember

  • STUBULATOR ROM hook (was: // Maybe instead of this, we could try requiring the STUBULATOR ROM?). The Stubulator is a small Atari-era PD ROM that bootstraps homebrew without needing the real BIOS. We removed the comment that suggested using it as a fallback. Worth considering as a v2.3 test-harness fixture (not shipping it as a user-facing fallback) — it's a small, known-state cart we can drop in for HLE-vs-real-BIOS A/B tests.
  • Old src/mmu.c removed in the source-tree reorganization. The original file emulated a Jaguar MMU (paging) that no shipping cart ever used; the only consumer was an Alpine debugger mode that's also been removed. If we ever add Jaguar dev-cart / networking / CD features that depend on memory protection, this is the file to resurrect (in master at the time of the reorg).
  • vjs.hardwareTypeAlpine field + alpineROMPath (master had these, branch removed them). Alpine was an Atari developer board with extra debug RAM at $C00000-$DFFFFF. Removing the option leaves RetroArch users without an "Alpine mode" toggle; if homebrew developers ever ask for it back, restore the field, the file-load path in src/core/file.c::JST_ALPINE, and the option entry in libretro_core_options.h.
  • NEW_SCOREBOARD #ifdef arms in src/jerry/dsp.c. Master had 19 references; branch has 15 (4 removed). The remaining #define NEW_SCOREBOARD toggles a hotter scoreboard path; the removed #else arms were the cold/legacy variant kept for a comparison that nobody runs anymore. No functional change, but if the new path ever regresses, the historical alternative is in master's src/dsp.c.

Comments / TODOs that should not fall off

  • src/tom/blitter.c !!! FIX !!! comments (DCOMPEN 8/16 BPP-only, 1-bit expansion behaviour). Pre-existing from upstream; wrap into the OP/blitter audit task already on this list.
  • HLE sound-engine auto-ack in src/jerry/dsp.c::DSPReadLong — current behaviour zero-acknowledges DSP-RAM polls in the DSP_SOUND_CMD_BASE..DSP_SOUND_CMD_END range when the DSP isn't running and we're in HLE. This is a workaround for games that poll sound-engine flags the BIOS engine would normally clear. Documented inline; revisit when the BIOS DSP audio engine is actually replicated (Wolf3D / Skyhammer / IS2 family).
  • JERRYI2SCallback SCLK/SMODE coupling (commented at top of src/jerry/dac.c). Output is hard-coded 48 kHz; SCLK changes affect DSP I2S rate but don't resample the host output. Future: add proper resampling so games that vary SCLK at runtime for pitch effects work.
  • test/baselines/{jagniccc,yarc}.png are 241 px tall (one row taller than the 240-line NTSC visible region). The extra row comes from miniretro in our regression harness — it captures one row past max_height for some video paths. Not a Virtual Jaguar bug; if we ever switch capture tools the baselines need re-shooting.

Code-organization items for v2.3.0

  • src/jerry/dsp.c is ~3000 lines with several long functions; split by responsibility (decode / dispatch / IRQ / scoreboard / state) so it's easier for humans + LLMs + static-analysis tools to reason about.
  • link.T export gating — done in v2.2.0. Production link.T now exports retro_* only; link-test.T carries the wider symbol surface for the white-box test harnesses (DSP*, dsp_, m68k_, GPU*, gpu_, JERRY, TOM*, OP*, Jaguar*, jaguarMainRAM, jagMemSpace, regs, sclk, smode, lowerField, vjs, ...). Makefile picks link-test.T when TEST_EXPORTS=1, which the test target sets automatically. Remaining for v2.3.0: macOS / iOS / tvOS dylibs ignore --version-script and currently still export everything with default visibility. Slim with -Wl,-exported_symbols_list,exports.list + a per-platform exports list, gated the same way (test builds get the wide list).
  • GIT_VERSION / CORE_VERSION are spread between Makefile and libretro.c. Move to a generated version.h so both sites read the same source of truth.
  • HLE constants block in src/core/jaguar.c was moved from inside JaguarReset() to file scope this PR — done. Future: consider promoting the BIOS-known register addresses into a shared header so other subsystems (e.g. test/test_hle_bios.c) don't have to hard-code them.
  • Optional: add static-analysis bots / linting / dead-code finders / const-correctness audits as a CI step. clang-tidy and cppcheck would be good starting points; the codebase already has a C89 lint, so the infrastructure is there.

Original docs/TODO items still relevant (Shamus / CJ)

The historical docs/TODO from the upstream Virtual Jaguar tree lists several still-open accuracy / feature items. These map onto the acid-test categories in test/acid/README.md; tracking here so they don't get lost:

  • "Fix VC behavior to match what a real Jaguar does. Still not sure just what the heck is going on there." [Shamus] — acid timing/. Active suspect for the Doom 1.5-2x speed regression (issue #131).
  • "Cycle accuracy for GPU/DSP/OP/Blitter." [Shamus] — cross-cutting; informs every category in test/acid/, especially bus/ (which can't pass without it).
  • "Need to propagate blitter fixes in the A1 <- A2 direction to the A1 -> A2 direction and the GPU fixes to various instructions to the DSP." [Shamus] — acid blitter/ (A1↔A2 symmetry tests) and gpu/ + dsp/ (shared opcode coverage).
  • "Blitter needs fixing." [Shamus] — acid blitter/. PR #129 fixed a perf-relevant chunk (ADDARRAY etc); accuracy axis still wide open.
  • "Need to emulate bus contention." [Shamus] — acid bus/. Almost certainly load-bearing for the Doom regression and the AvP audio dropouts.
  • "Need to fix timing in the OP. As it is now, it gives a false impression of how much it's capable of." [Shamus] — acid op/.

The original docs/TODO is intentionally left untouched — it's the authors' historical record and we track our own work via GitHub issues + this file + test/acid/.