libretro/virtualjaguar-libretro

0 Fork 0

mirror of https://github.com/libretro/virtualjaguar-libretro.git synced 2026-11-07 04:36:00 -10:00

Files

T

Joseph Mattiello f6e2e7b784 docs(release): refresh final-core cartridge and CD boot matrices

2026-10-20 16:55:25 -04:00

90 KiB

Raw Blame History

CD boot matrix

Generated by test/tools/cd_boot_matrix.sh -- the regression gate for CD-boot work. Every subsequent CD-boot fix must move rows forward through the stage taxonomy below, never backward. Do not hand-edit this table; re-run the script and commit its output instead. test/tools/matrix_diff.py OLD.md NEW.md enforces the never-backward rule against the previous tag's copy (checklist: docs/release-process.md).

How to re-run

bash test/tools/cd_boot_matrix.sh

Requires test/roms/private/ populated with the CUE/CDI/ISO images and the real Jaguar CD BIOS (Jaguar CD BIOS.rom in this repo's corpus; see test/test_cd_bios_boot.c's header comment for the full accepted-filename list). In a git-worktree checkout that does not share untracked files with the main checkout: ln -s /path/to/main-checkout/test/roms/private test/roms/private. Core must be built make TEST_EXPORTS=1 (a plain make strips the dlsym export set the harnesses need -- the script auto-rebuilds if it detects this). Env knobs: CD_MATRIX_FRAMES (default 3000), CD_MATRIX_TIMEOUT (default 120s per title x mode), CD_MATRIX_ROMS_ROOT (default test/roms/private), CD_MATRIX_OUT (output path, default docs/cd-boot-matrix.md), CD_MATRIX_LOGDIR (raw per-run logs, default a fresh mktemp -d; pass a fixed dir to keep logs across chunked invocations), and CD_MATRIX_MAX_RUNS (max NEW runs per invocation, default 0 = unlimited).

Rows append incrementally with a resume guard: already-recorded title x mode combos are skipped, so an interrupted sweep resumes where it left off. For long sweeps (or agent sessions with per-command time limits), run chunked:

# ~3 runs x <=120s wall each per invocation; repeat until every line
# reports "already recorded, skipping":
CD_MATRIX_MAX_RUNS=3 CD_MATRIX_LOGDIR=/tmp/cdmx_logs bash test/tools/cd_boot_matrix.sh

To force a re-run of specific rows, delete those table lines from the output file first -- the resume guard re-runs only missing title x mode combos.

Legend

Stage (finest level the harness evidence supports; ? = insufficient evidence to classify further, never fabricated):

Stage	Meaning
`LOAD_FAIL`	`retro_load_game()` returned false, or the harness child crashed (segfault) before producing a usable PC trace
`BIOS_INTRO`	Real CD BIOS executing (boot ROM cube animation, BIOS main loop/error handler, or CD-read poll code) -- has not yet handed off to the boot stub or game
`BOOT_STUB`	PC parked in the injected boot-stub ISR/poll-loop/data region (`$080000`-`$08FFFF`)
`GAME_CODE`	PC in game-owned RAM outside the BIOS/boot-stub bands (HLE: any stable multi-PC run, since HLE synthesizes past BIOS/boot-stub by design)
`MENU` / `IN_GAME`	Not distinguished by any current harness -- see "Headless framebuffer caveat" in CLAUDE.md. Reported as `GAME_CODE` with a note.
`HARNESS_HANG`	Killed at the wall-clock timeout; the harness process itself did not exit
`? (pc_escape)`	Final 68K PC outside every valid execute window (RAM <`$200000`, boot ROM `$E00000`-`$E1FFFF`, cart/CD BIOS `$800000`-`$8FFFFF`) -- a crash, not a reached stage
`? (BIOS service band ...)`	Final PC in `$0`-`$3FFF` (BIOS TOC / CD_read service code, used both by the BIOS during boot and by games after handoff). Two variants, distinct evidence strength: "PC-SET showed no game-band entries" = the visited-PC dump was checked and contains no `$090000`-`$1FFFFF` addresses (checked negative); "PC-SET suppressed, unique_pcs>32" = the harness never emitted the dump, so the game-band check could not run (undetermined)

Score N/1 follows the harness's own PASS/FAIL heuristic (PC stays in RAM/BIOS/cart range, not self-looping, not thrashing among <=4 distinct PCs, RAM has non-trivial non-zero payload) -- see test/cd_assertions.h. It is informational, not the stage classification: a title can score FAIL yet still classify as GAME_CODE (e.g. stuck in its own hardware-poll wait loop after booting), or score PASS while stuck in a still-early BIOS/boot-stub band.

Baseline notes -- 2026-07-13, branch feature/jaguar-cd-support @ `cbd2518`

Dated observations for this baseline run (3000 frames/title/mode, 120 s wall-clock cap). Notes sections are snapshots -- a regenerated table gets a fresh dated section; do not retrofit old ones.

Stage distribution (21 rows): HLE -- 9 GAME_CODE (of which 3 score 0/1: Baldies, Iron Soldier 2, and Baldies' final PC $1FFFDE is top-of-RAM/stack, i.e. GAME_CODE only in the "68K executing game-owned RAM" sense), 1 LOAD_FAIL (baldies.cdi segfault). BIOS -- 4 BIOS_INTRO, 1 BOOT_STUB (Primal Rage, final_pc=$0803AC = documented boot-stub poll loop), 1 GAME_CODE (BrainDead 13 -- the disambiguating game-band PCs $124342 $124346 $12434C are quoted in the row's own evidence column), 2 ? (BIOS service band) (Highlander: PC-SET checked, no game-band entries; Iron Soldier 2: PC-SET suppressed at unique_pcs=34, so the game-band check never ran -- undetermined, not a checked negative), 1 ? (pc_escape) (Battle Morph, final_pc=$8FBFB758), 1 HARNESS_HANG (Baldies, video_stall watchdog fired at frame 745, GPU wedged at $F03060), 1 LOAD_FAIL (baldies.cdi), plus the ISO documentation row (BIOS_INTRO, final_pc=$194D12 = CD BIOS poll loop; .iso support has since been removed entirely (load refused with an explanatory error) — historical note only; it was a permanent, documented limitation -- no session-2 layout).
No title reaches a verified MENU or IN_GAME headlessly. Every GAME_CODE row still trips video_stall or sits in a tight wait loop. MENU/IN_GAME verification requires RetroArch (headless framebuffer caveat).
BIOS-mode error-loop signature: final_pc=$0059B0 with unique_pcs=6 and seeks=0 (seen in shorter runs / Task 2) is the BIOS "no boot" spin. At 3000 frames several titles progress past it: Dragon's Lair and Space Ace show real disc activity (seeks=1, dsaIRQs in the millions, cdPlaying=1) yet remain in BIOS code bands -- boot progresses but never completes.
Task 2 aggregate parity: harness scores here are per-title at 3000 frames and are NOT comparable 1:1 to the 300/900-frame Task 2 sweeps (test_cd_hle_boot 6/9, test_cd_bios_boot 7/9); longer runs legitimately change PASS/FAIL for titles that wedge late (e.g. Battle Morph HLE passes at 3000 frames, unique_pcs=126). The stage column, not the score, is the regression gate.
baldies.cdi segfaults the core in both modes (CDI parser crash, known since the corpus was assembled -- see test/cd_assertions.h cd_ext_enabled() rationale). Fixing the CDI loader is its own work item; the row exists so the fix has a baseline to move.
Last-DSA-command column from the original plan sketch is intentionally absent: per-title DSA tracing is Task 4 instrumentation. The [CD-DIAG] counters (seeks, dsaIRQs, fifoReads) in the raw logs carry the equivalent evidence for now.

Results

Title	Mode	Score	Stage	Watchdog	PC evidence
Baldies (USA) (Rev 1).cue	hle	1/1	GAME_CODE	(none)	[PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20409B unique_pcs=39 final_pc=$05FE8A
Baldies (USA) (Rev 1).cue	bios	1/1	GAME_CODE	(none)	[PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20303B unique_pcs=65 final_pc=$05FE88
Battle Morph (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Battle Morph (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=26709B unique_pcs=47 final_pc=$007794
Battle Morph (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Battle Morph (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=28474B unique_pcs=69 final_pc=$022052
BrainDead 13 (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=17926B unique_pcs=51 final_pc=$004FD0
BrainDead 13 (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=19577B unique_pcs=62 final_pc=$004FD2
Dragon's Lair (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=18477B unique_pcs=74 final_pc=$00967A
Dragon's Lair (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20802B unique_pcs=79 final_pc=$004C10
Highlander - The Last of the MacLeods (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=21068B unique_pcs=52 final_pc=$009222
Highlander - The Last of the MacLeods (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=16336B unique_pcs=69 final_pc=$0088DA
Hover Strike - Unconquered Lands (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=6150B unique_pcs=48 final_pc=$05DAA4
Hover Strike - Unconquered Lands (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=15762B unique_pcs=74 final_pc=$05DAA6
Iron Soldier 2 (USA) (Songbird).cue	hle	1/1	GAME_CODE	(none)	[PASS] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29325B unique_pcs=23 final_pc=$00B542
Iron Soldier 2 (USA) (Songbird).cue	bios	1/1	GAME_CODE	(none)	[PASS] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29752B unique_pcs=41 final_pc=$00B542
Primal Rage (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=28237B unique_pcs=97 final_pc=$00457A
Primal Rage (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=25774B unique_pcs=75 final_pc=$01AFCE
Space Ace (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=19041B unique_pcs=60 final_pc=$004D4E
Space Ace (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20881B unique_pcs=78 final_pc=$00B87A
baldies.cdi	hle	1/1	GAME_CODE	(none)	[PASS] baldies.cdi : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20409B unique_pcs=39 final_pc=$05FE8A
baldies.cdi	bios	1/1	GAME_CODE	(none)	[PASS] baldies.cdi : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20303B unique_pcs=65 final_pc=$05FE88
Myst (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Myst (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=10988B unique_pcs=256+ final_pc=$0103FA
Myst (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Myst (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=8123B unique_pcs=256+ final_pc=$00EF82
Baldies (USA) (Rev 1).chd	hle	1/1	GAME_CODE	(none)	[PASS] Baldies (USA) (Rev 1).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20409B unique_pcs=39 final_pc=$05FE8A
Baldies (USA) (Rev 1).chd	bios	1/1	GAME_CODE	(none)	[PASS] Baldies (USA) (Rev 1).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=20303B unique_pcs=65 final_pc=$05FE88
Myst (USA).chd	hle	1/1	GAME_CODE	(none)	[PASS] Myst (USA).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=10988B unique_pcs=256+ final_pc=$0103FA
Myst (USA).chd	bios	1/1	GAME_CODE	(none)	[PASS] Myst (USA).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=8123B unique_pcs=256+ final_pc=$00EF82
Frog Feast (USA) (Unl).chd	hle	1/1	GAME_CODE	(none)	[PASS] Frog Feast (USA) (Unl).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=7027B unique_pcs=119 final_pc=$006E92
Frog Feast (USA) (Unl).chd	bios	1/1	GAME_CODE	(none)	[PASS] Frog Feast (USA) (Unl).chd : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=9667B unique_pcs=131 final_pc=$006E92

Raw per-run logs are not committed; re-run the script to regenerate.

Regenerated on `ab77d16` -- 2026-10-07 (v3.7.1 release gate)

Swept with the v3.7.0 tag (274fa74) and libretro/develop (ab77d16) built side by side from pristine trees (CD_MATRIX_TIMEOUT=500, 3000 frames, the host heavily loaded). The two builds agree on all 28 rows: every row is 1/1 GAME_CODE, no watchdog line, and final_pc / unique_pcs match for every row where the harness prints them. No CD regression since v3.7.0.

Against the rows this table carried before (4cbb1d56cff7, the 2026-10-02 sweep), the only change is that the five bios rows that read 0/1 now read 1/1 with no cd_seek_wedge line: Baldies (USA) (Rev 1) .cue / .chd / .cdi and Myst (USA) .cue / .chd. The v3.7.0 tag already reads that way, so it predates every PR merged since and v3.7.1 adds nothing here. The bios rows of Battle Morph, BrainDead 13, Highlander and Primal Rage, which carried a cd_seek_wedge line in the old table, no longer do either (same reason). ram_payload and unique_pcs moved slightly on a few rows (for example Frog Feast unique_pcs 122 -> 119, Hover Strike hle ram_payload 13546B -> 6150B) between the old table and v3.7.0; they are identical between v3.7.0 and ab77d16.

Regenerated on `2f913d8` -- 2026-10-02 (#735)

The first full sweep since 2026-08-13. Against those rows, two titles improved and four regressed. Everything else is unchanged at the stage and verdict level.

Improved. Frog Feast (USA) (Unl).chd: hle LOAD_FAIL -> GAME_CODE PASS, bios BIOS_INTRO -> GAME_CODE PASS.
Regressed, all in bios mode, PASS -> FAIL at the same stage. Baldies .cue / .chd / .cdi and Myst (USA).cue. In each the 68K is parked waiting for an interrupt that does not come (not_loop=0). These are #752 (Baldies: STOP at $05F4EA, GPU completion IRQ starved) and #756 (Myst: 68K sleeps on STOP, GPU IRQ serviced early). #757 claims #756 and should move these rows.

Wall-clock caps were raised for this sweep (CD_MATRIX_TIMEOUT=400) because the host was heavily loaded. The frame budget is unchanged (3000).

Myst CHD joined the sweep -- 2026-10-01

Myst (USA).chd now has rows (issue #774, after #755 fixed virtual-pregap handling in ParseCHD). It is the corpus's only CHD with virtual pregaps: tracks 3-10 carry 148-149-frame VAUDIO pregaps, which is exactly the case #755 fixed. Its rows match the Myst (USA).cue rows field for field when both are measured on the same build (ram_payload, unique_pcs, final_pc, and in bios mode the same cd_seek_wedge frame, seek and drain counts), so the CHD reader returns the same data the CUE/BIN path does.

Conversion matters. Homebrew/distro chdman (0.289 here) does not write CHSE session tags, and the core refuses such a CHD at load (LOAD_FAIL). Convert with tools/jagcd/jagcd-to-chd and the pinned chdman from the jagcd-tools-*.zip release asset. It exits 2 on this disc, which means "virtual pregaps present": a warning, not a failure.

These two rows are stamped with a newer cache id than the rest of the table. The 2026-08-13 rows are stale (#735). On the current build the Myst CUE rows also read not_loop=0, and bios reads [FAIL]. That is identical on v3.6.1 (f9a3c89), so it predates #750 and #755. Whether it is a regression or Myst idling on a static screen at frame 3000 is for #735 to decide.

Myst joined the sweep -- 2026-07-30

Myst (USA).cue was missing from this script's CUE_TITLES array even though cd_discover_discs() has always found it in the corpus -- the matrix simply never asked for it. It now has rows: GAME_CODE 1/1 in both hle and bios, matching the independent evidence already on record (cd_wedge_probe runs clean in both modes; cd_visual_verify gives avg RMS 2548 bios / 2698 hle, both in envelope).

Two things worth recording, because neither matches the assumption that prompted the addition:

The $004000-$007FFF band gate is not what makes Myst pass. Myst's boot executable does load at $005000-$01D380 (src/cd/jagcd_bios.c:45), inside the shared band, so it looked like a fourth title the pre-5a83d3d unconditional rule would have mislabelled BIOS_INTRO. It is not: at 3000 frames the sampled PC is already well past that code -- final_pc=$00CA64 (hle), $00F1A4 (bios), both outside the band. Sampling Myst much earlier would be needed to land in it.
The bios row's cd_seek_wedge line was a false positive (fixed in #741): Myst goes CD-idle for ~6s during the intro movie's all-black pause, with the transfer finished (BUTCH bit 0 cleared). The watchdog no longer counts that window. The hle row shows the same window as video_stall. Neither is a failure.

Adding a title also exposed a latent bug in this script, fixed in the same commit: new rows were appended to EOF instead of into the results table. On a fresh generation the table is the last thing in the file so it made no difference, but on a resume against this doc -- which carries several hundred lines of prose notes after the table -- a new title's rows landed outside the block find_row_lineno scans. The guard could never see them, so the title re-ran on every invocation and appended another duplicate each time. Only adding a title to CUE_TITLES can trigger it, which is why it survived this long.

`baldies.cdi` must be regenerated with `cue2cdi`, not taken from the ROM set

Recorded because it cost a full sweep to work out, and the next person to rebuild the corpus will hit it again.

The corpus's baldies.cdi is not the baldies.cdi that ships with the ROM set. It is generated from the CUE/BIN by test/tools/cue2cdi, which emits the layout src/cd/cdintf.c :: ParseCDI walks natively. When test/roms/private was rebuilt from a separate archive during this work the ROM-set download was substituted, and both rows went backward -- hle GAME_CODE -> LOAD_FAIL, bios GAME_CODE -> BIOS_INTRO. The two images differ in size (61.7 MB generated vs 62.4 MB downloaded) and place track 2 at different offsets ($AB0270 vs $B06490); at the download's offset the boot-stub reader finds zeros, logs [CD-BOOTSTUB] Boot stub extraction failed, and the HLE load is refused.

Regenerating restores both rows exactly -- evidence strings identical to the previous sweep, down to ram_payload=18235B unique_pcs=29 final_pc=$05FE82 (hle) and the whole of the bios row's cd_seek_wedge line:

./test/tools/cue2cdi \
  "$ROMS/Baldies (USA) (Rev 1)/Baldies (USA) (Rev 1)/Baldies (USA) (Rev 1).cue" \
  "$ROMS/baldies.cdi" --verify

This is not evidence that the downloaded image is bad. It is a legitimate DiscJuggler image that our ParseCDI boot-stub path cannot read -- a real gap, independent of every row in this table, and not tracked anywhere yet.

Re-run notes -- 2026-07-29, branch fix/cd-fmv-bios-handoff @ `17375c3`

Full re-sweep (3000 frames, CD_MATRIX_TIMEOUT=400, chunked via CD_MATRIX_MAX_RUNS) after correcting the $004000-$007FFF stage rule in classify_stage(). Three rows moved forward, zero backward -- and no emulator code changed, because there was no emulator defect:

BrainDead 13 bios: BIOS_INTRO -> GAME_CODE (final_pc=$004FCA)
Dragon's Lair bios: BIOS_INTRO -> GAME_CODE (final_pc=$004C12)
Primal Rage bios: BIOS_INTRO -> GAME_CODE (final_pc=$0044CC)

Every other row is unchanged in stage and score (all 20 rows 1/1 GAME_CODE in both modes). The at-risk bios rows the taxonomy change could have disturbed -- Baldies, Battle Morph, Highlander, Hover Strike, Iron Soldier 2, Space Ace -- all still report GAME_CODE 1/1. final_pc and unique_pcs vary run to run inside GAME_CODE (the harness samples a live game); that is not a stage move.

Why these three rows were never stuck

docs/cd-known-issues.md item 1 recorded these as a real "FMV titles do not hand off from the real BIOS" defect, with three ranked timing candidates in docs/cd-bizhawk-comparison.md. It was none of them: the band rule was wrong, and the rows were false negatives.

Method (the comparison doc's own prescribed first step -- disassemble the loop instead of correlating against it). cd_wedge_probe --arm N --freeze-frames 0 forces a full state dump at a chosen frame, and the snapshot was disassembled at each reported PC:

Title	PC	What the code is	Polled address	Written by
Dragon's Lair	`$004C0A`-`$004C12`	`move.l $562E,d0 / cmp.l d0,d2 / bgt`	`$562E`	GPU PIT ISR (`$F0308A`), 30 Hz
BrainDead 13	`$004FCA`-`$004FD2`	byte-identical code, `$72DA` instead	`$72DA`	GPU PIT ISR, 24 Hz
Primal Rage	`$004436`-`$0044EE`	joypad scan: writes `$817F/$81BF/$81DF/$81EF` to `$F14000`, de-interleaves via tables at `$44F0/$4510/$4530`	-- (not a poll loop)	--

Dragon's Lair and BrainDead 13 are both ReadySoft titles and share one FMV player, linked at different addresses -- the instruction stream at $004BF0 and $004FB0 is byte-for-byte identical. The loop waits for the movie's presentation clock to reach the next frame's timestamp. That clock is a 48-bit fixed-point accumulator ($562E.w integer : $5630.l fraction) that the GPU's PIT ISR advances by a per-title increment; it was measured still advancing across snapshots at frames 1200/1800/2400/2900 (DL: $8E -> $1BA -> $2E6 -> $3DF, exactly 30.0/s; BD13: $CD -> $1BD -> $2AC -> $374, 24.0/s). Sampling landed mid-wait for the next 1/30 s tick. Primal Rage was not even in a loop -- its PC ring is a straight-line sweep through its input scanner.

Confirmed with cd_visual_verify in bios mode (screenshots read, not just counted): Dragon's Lair plays its attract FMV (79.5% non-black, avg RMS 1230); BrainDead 13 plays its castle-and-moon intro (68.1%, RMS 2651); Primal Rage runs its full attract cycle -- "WHO WILL RULE THE NEW URTH?", then DEMO gameplay with two dinosaurs fighting, then back to the "JAGUAR VERSION 1.2" title screen (93.3%, RMS 980). Primal Rage needed 6000 frames to show the cycle: its zero-motion windows (3060-3359, 5280-5519) are black scene transitions, the same benign pattern already documented for Myst, not a freeze.

The decisive datum needs no visuals at all: Dragon's Lair's hle and bios rows now both report final_pc=$004C12, and BrainDead 13's both report $004FCA -- the same instruction in the same player. Only the bios path applied the band rule, so the same code was labelled GAME_CODE in one mode and BIOS_INTRO in the other. Space Ace vs Dragon's Lair (the control pair the comparison doc built its ranking on) was likewise never "works vs stuck": Space Ace's copy of the same ReadySoft player is linked at $00B87A, outside the band.

Re-run notes -- 2026-07-27, branch feature/jaguar-cd-support @ `73438ba`

Full re-sweep (3000 frames, CD_MATRIX_TIMEOUT=300) after four CD fixes. Five rows moved forward, zero backward:

Battle Morph bios: ? (pc_escape) -> GAME_CODE 1/1 -- 8c9356b (HLE stream status-struct base latched per CD_read; stops the mid-stream ISR-setup shred of the game's uploaded GPU worker).
Iron Soldier 2 hle: 0/1 -> 1/1 GAME_CODE -- 73438ba (game-installed CD drivers own the transfer; HLE hooks step aside when their $4E75 stubs are overwritten) on top of bf4b7e8 (G_CTRL SINGLE_STEP honored, which fixed the deeper match-load wedge in bios mode).
Baldies hle: 0/1 -> 1/1 GAME_CODE -- 73438ba (FIFO interlock + 32-byte-batch progress cadence).
baldies.cdi hle + bios: LOAD_FAIL (segv) -> GAME_CODE 1/1 both -- 6fc874c/c5707c8 (canonical packed DiscJuggler track-block layout in ParseCDI + single-file session-2 boot-stub extraction). NOTE: the on-disk baldies.cdi asset was regenerated with the fixed cue2cdi (9c82be1); files written by the pre-9c82be1 writer use a mirrored, non-canonical layout and are invalid -- reconvert any old .cdi. Final PCs now match the .cue rows exactly in both modes (CDI parity).

One transient during the sweep: Battle Morph bios first classified HARNESS_HANG at the default 120s wall-clock while other heavy test sweeps shared the host; re-run at CD_MATRIX_TIMEOUT=300 passes. Prefer the longer timeout when the host is loaded.

Re-run notes -- 2026-07-13, Task 6A fix (BUTCH -> GPU IRQ1 routing)

Five bios rows deleted + re-run after the Mechanism-A fix (3000 frames, same knobs); all other rows are the earlier baseline.

Targets -- Primal Rage / Highlander / Iron Soldier 2 (bios): the FIFO-never-fills deadlock is gone. Primal Rage's watchdog line now shows gpu_pc=$00F03010 (executing the CD ISR; was foreground poll $F03060), and a VJ_CD_TRACE=1 run shows a live FIFO_FILL/FIFO_DRAIN stream with the block counter advancing (116618 -> 118131 in 1500 frames; dsaIRQs 556791 -> 1, i2sEn=1). Highlander now records fifo_drains=54 (was 0). Stage labels have not advanced yet (Primal Rage still BOOT_STUB even at a 12000-frame probe, 9.7M FIFO reads): the boot stub's sentinel scan never matches despite real word-swapped sector data flowing -- that is a separate downstream mechanism (data content/positioning), not the Mechanism-A gate.
Controls -- BrainDead 13 (bios) and Dragon's Lair (bios): stage, score, and PC evidence identical to baseline (Dragon's Lair row is byte-identical; BrainDead differs only in which watchdog signature got logged first -- cd_seek_wedge vs video_stall, both fire at the same frame, and its stalled gpu_pc moved $F03270 -> $F03276 within the same GPU poll band; its game-band PC set $124342 $124346 $12434C is unchanged). No regression.
HLE rows re-validated post-reroute (BUTCHExec ticks for all bootConfig.isCDGame modes, so the IRQ0->IRQ1 change is live in HLE too): Primal Rage (hle), Iron Soldier 2 (hle), BrainDead 13 (hle) deleted + re-run -- all three unchanged vs baseline (Primal Rage and IS2 rows byte-identical; BrainDead identical stage/score/final_pc, the re-run row merely adds the informational PC-SET game-band list). No spurious $F03010 dispatches surfaced in HLE mode.
Provenance: the reroute functionally reverts commit 3279b30 ("CD: fix BUTCH->GPU IRQ line - route to GPU IRQ0 (EXT1), not IRQ1 (DSP)", 2026-04-29), whose "$F03000 = EXT1" premise was unverified and which self-declared "no behaviour change without the BUTCHExec scheduler tick". The original IRQ1 routing it removed was correct.

Re-run notes -- 2026-07-14, Task 6D fix ($2C00 TOC track-indexed layout)

Rewrote the $2C00 TOC injector (both src/cd/jagcd_bios.c and the HLE path's HLEPopulateTOC in src/cd/jagcd_hle.c) to the authoritative track-indexed layout produced by the real CD BIOS's own $2C00 TOC builder (disasm ROM $808BE8: entry for track N at $2C00 + N*8, byte[0]=track#, byte[1..3]=start MSF, byte[4]=0-based session). The old standalone zero-longword marker slot terminated Baldies' $4E18 scan early (deliberate ILLEGAL) and left every byte[4] zero so the session-key match was impossible; both defects fixed. The derivation is captured by the RED/GREEN contract test test/test_cd_toc_contract.c. Boot-mode determination: outcome (c) — $2C00 is PRNG garbage at injection time (the BIOS TOC builder never runs under the auth-bypass boot), so the injected table must be correct itself.

Writer provenance (reconciling prior project notes): $808BE8 is a 68K routine in the BIOS ROM — every quoted instruction is 68K (movem.l/lea/dbra) and it polls BUTCH DSA responses directly at $DFFF0A — not the DSP code the prior notes said the BIOS copies to $F1B000 to query the drive via BUTCH $14xx TOC requests. The two are most plausibly different stages of one pipeline (DSP handles drive-level transport; this 68K routine consumes the $60-$64 DSA response words and builds $2C00), but this disasm does not confirm the DSP stage either way — the earlier "the DSP writes $2C00" belief was at minimum imprecise about which processor does the writing. The layout's real validation anchors are (1) the two independently disassembled game boot-stub scanners, which read exactly this track-indexed format, and (2) the reviewer-reproduced GREEN on the real Baldies disc. Note also that track-indexed and the old sequential placement coincide byte-for-byte when track numbering is contiguous from 1 (as on real discs) — the load-bearing fixes are the removed zero-longword marker slot and the byte[4] session stamp, not the indexing scheme itself.

Contract-test scope: test_cd_toc_contract's Primal Rage MSF assertion compares the injected table against CDIntfGetTrackInfo() — the injector's own data source — so it proves copy fidelity into the scanner-visible layout, NOT seek-target/LBA correctness (an H1-class pregap/offset bug would still pass it); Primal Rage's marker-then-NEXT-entry path is likewise not yet exercised end-to-end in-game (title blocked downstream of the TOC scan).

Target that moved -- Baldies (bios): HARNESS_HANG (frame-437 ILLEGAL halt; the boot stub's TOC scan returned -1) -> [PASS] final_pc=$003610, now in the BIOS CD-read poll band with the CD stream live (seeks=1, i2sEn=1, fifoReads=2.1M) -- the same downstream wall as Highlander/IS2. The ILLEGAL halt is cleared.
Targets unchanged (different blockers) -- Primal Rage / Battle Morph (bios): Primal Rage still BOOT_STUB $0803AC (its blocker is the downstream sentinel scan per Task 6A, not the TOC; the TOC is now faithfully track-indexed and its $0803E2 scanner lands on the correct track). Battle Morph still pc-escapes at frame 437 ($8CA88044) -- the Task 6B UNCONFIRMED "derailment <-> TOC" link is now evidence-negative: fixing the TOC did not change it; its derailment is a distinct, still-uncharacterized mechanism.
HLE rows (targets -- HLE injector also fixed): Primal Rage (hle) [PASS] final_pc=$00419E and Iron Soldier 2 (hle) final_pc=$007416 uniq=3 -- both byte-identical to baseline (their blockers are not the TOC scan; no regression).
Controls -- BrainDead 13 / Dragon's Lair / Highlander (bios): byte-identical to baseline (final_pc $00361E with game-band $124342/46/4C; $0060D8; $003610 respectively). No regression.
make TEST_EXPORTS=1 test exit 0 (56103/56103); c89-lint clean on both touched src files. New run-gated harness test/test_cd_toc_contract.c (needs VJ_TOC_DISC).

Diagnosis (Task 5) -- 2026-07-13, branch feature/jaguar-cd-support @ `5827940`

Trace-based root-cause analysis of the BIOS-mode boot failures, using the Task 4 instrumentation (VJ_CD_TRACE=1, test/test_cd_bios_boot / test_cd_hle_boot, 1500 frames). Every classification below cites trace lines, [CD-DIAG] counters, register values, or disassembly. Temporary per-register counters ([T5-DIAG]) were added to src/cd/cdrom.c for the run, cited here, and reverted -- they are NOT in the tree. Two distinct mechanisms account for all four failure classes.

Mechanism A -- I2S FIFO never enabled after seek (FIFO-wait deadlock)

RESOLVED (Task 6A, 2026-07-13). Attribution is now CONFIRMED: our emulation asserted the wrong GPU IRQ line. BUTCHExec raised GPU IRQ0 (CPU-sourced, vector $F03000) but the CD BIOS installs its CD-data ISR at $F03010 = the GPU IRQ 1 vector (DSP/JERRY source, JTRM vector = int# x 16), enables only G_FLAGS INT_ENA1 ($20), acks via INT_CLR1 and re-arms the JERRY external latch (J_INT=$0101) in its epilogue -- BUTCH's eint enters through JERRY (butch.v line 83). With only INT_ENA1 set, our IRQ0 latch was permanently masked: instrumented run showed zero GPU interrupt dispatches ever (enable-bail x 2.4M, latch=$01 vs enable=$02), zero GPU BUTCH reads, zero I2CNTRL writes. The ISR's bit-2 setter (BSET #2 at ISR +$4A on the BUTCH-bit-13 branch) simply never ran. Fix: GPUSetIRQLine(GPUIRQ_DSP, ...) in BUTCHExec. Post-fix: dsaIRQs 556791 -> 1 (response consumed on first dispatch), i2sEn=1, FIFO fill/drain engine live (111k cycles / 1500 frames). Remaining blocker for these titles is downstream (boot-stub sentinel scan never matches despite real word-swapped sector data flowing -- a data-content/positioning mechanism, tracked separately).

Affects: Primal Rage (bios) [class 1, BOOT_STUB $0803AC] and Highlander (bios) [class 3]; Iron Soldier 2 (bios) [class 3] shares the same deadlock with a partially-unresolved chain (see IS2 caveat below).

Verdict is split in two, at different confidence levels:

The deadlock itself -- CONFIRMED (airtight): I2CNTRL bit 2 (i2s_fifo_enabled) is never set by any processor after the seek, so BUTCHExec's bit-2 FIFO-fill gate never opens, FIFO_FILL never fires, and the guest polls forever.
Attribution (whose fault) -- UNCONFIRMED: whether the guest never intends to set bit 2 on this path, or a guest GPU CD-ISR is supposed to set it but our GPU/IRQ emulation stalls it before it gets there, is not resolved by this evidence (see below).

The DSA/seek handshake completes correctly. Primal Rage ring tail:

#7 DSA_TX  $1244   #8 SEEK_START $1244 block=116618   #9 SEEK_DONE $0100 block=116618
[CD-DIAG] butchExec=786000 globalDisabled=229109 seeks=1 fifoIRQs=0 dsaIRQs=556791 fifoReads=2 cdPlaying=1 fifoReady=0 i2sEn=0

SEEK_DONE fires 100 ticks after SEEK_START (== SEEK_DELAY_TICKS), the $0100 "Found" response is queued -- the seek is NOT the problem. But no FIFO_FILL event ever appears and i2sEn=0.

Root cause, confirmed by temporary per-register counters:

Primal Rage  [T5-DIAG] dsDataReads=3 (afterSeek=0) i2cntrlWrites=6 (afterSeek=0) lastI2cWrite=$0001@PC$00365E i2cReg=$00000001
Highlander   [T5-DIAG] dsDataReads=3 (afterSeek=0) i2cntrlWrites=6 (afterSeek=0) lastI2cWrite=$0001@PC$00365E i2cReg=$00000001
Iron Sold 2  [T5-DIAG] dsDataReads=8 (afterSeek=4) i2cntrlReads=5804936 i2cntrlWrites=10 (afterSeek=4) lastI2cWrite=$0011@PC$00365E i2cReg=$00100011

The CD BIOS CD-read routine (PC $00365E) writes I2CNTRL = $0001 (bit 0 = i2s_drive, per test/mister_ground_truth.h butch.v map). It never sets bit 2 (i2s_fifo_enabled, mask $04) -- for Iron Soldier 2 the last write is $0011 (bits 0+4; bit 4 is the read-only FIFO-not-empty status line, not the enable). i2cReg (live register bytes) confirms bit 2 clear in every case.
src/cd/cdrom.c BUTCHExec() gates FIFO fill on I2CNTRL bit 2 (cdRam[I2CNTRL+3] & 0x04, lines ~595 and ~622). With bit 2 never set, the delayed-fill countdown perpetually re-arms (fifoFillDelay = 1; // Retry next tick) and fifoDataReady never becomes true. FIFO never fills; the guest polls forever (Primal Rage's boot stub at $0803AC; Highlander/IS2 in the BIOS CD_read poll band $003610-$0036BE, which longword-reads $00DFFF00 = BUTCH status).

Iron Soldier 2 caveat (confirmed root, scoped-UNCONFIRMED chain): the shared confirmed fact -- I2CNTRL bit 2 never set, i2sEn=0, fifoReady=0 -- holds for IS2. But IS2's exact deadlock chain is not identical to Primal Rage's: IS2 reads DS_DATA 4 times after the seek (dsDataReads afterSeek=4), and the DS_DATA $0100-pop branch (cdrom.c ~763-773) sets fifoDataReady=true unconditionally (independent of bit 2). Yet IS2 still ends fifoReady=0 with 5.8M FIFO_DATA reads and fifo_drains=0. Whether IS2 pops $0100 at all and, if so, what re-clears fifoDataReady before a drain completes is not fully resolved (would need a one-run fifoDataReady- transition trace). So IS2 = same confirmed deadlock (bit 2 never set), with its FIFO drain/refill interaction left UNCONFIRMED; the Primal Rage / Highlander deadlock chain below is the airtight one.

What the write counters do and do not prove. The [T5-DIAG] counters lived in the common CDROMWriteWord/CDROMWriteByte handlers -- the single choke point every processor's CD-range writes funnel through (src/core/jaguar.c:480,623 for the 68K path, :447,592 for the who-tagged GPU/DSP/DMA path) -- so "no bit-2 write" holds for all processors, and no write was mis-decoded or dropped (the values that were written, $0001/$0011, are stored correctly; i2cReg reflects them). But that only proves no processor executed a bit-2 write, NOT that the guest never intended one. The CD-data path is architecturally driven by a guest GPU CD-ISR that the CD BIOS uploads into GPU RAM -- src/core/jaguar.c:762 asserts GPU IRQ0 precisely "to drive the CD-data ISR". The trace signature here -- 556k GPU IRQ0 assertions, GPU running (gpu_run=1, PC circling $F03060-$F03080 in GPU RAM), yet dsDataReads afterSeek=0 and zero post-seek I2CNTRL writes -- is equally consistent with that guest ISR firing but never completing its DSA-consume -> I2CNTRL handshake, i.e. the "our GPU/IRQ emulation breaks the path" branch. Attribution between "guest never issues" and "guest ISR stalls under our GPU/IRQ emulation" is therefore UNCONFIRMED.

Related confirmed symptom (cause vs effect unresolved, same ambiguity): dsDataReads afterSeek=0 for Primal Rage/Highlander means the $0100 seek response is never consumed via DS_DATA. dsaResponseReady stays latched true, so BUTCHExec re-asserts GPU IRQ0 on every interrupt-enabled halfline (dsaIRQs=556791) -- the DSCNTRL-read latch-clear at cdrom.c:743 is on a path that never runs here. If the guest GPU ISR is supposed to consume the response and set bit 2, this un-consumed response and the IRQ storm are the same stall observed from the DSA side, not a separate fault.

Task 6 pointers (hardware-correctness is Task 6's call), three candidates to resolve against JTRM/MiSTer butch.v -- in this order:

Determine whether the CD BIOS's uploaded GPU IRQ0 CD-ISR is expected to consume the DSA response and set I2CNTRL bit 2 -- and if so, why our GPU execution of it stalls (disassemble the GPU-RAM ISR the BIOS uploads; check its progress against the GPU PC band $F03060-$F03080 observed wedged). This MUST be resolved first: if the ISR is being blocked by a GPU/IRQ emulation bug, any change to the FIFO gate would mask it.
Check MiSTer butch.v FIFO-fill logic: does hardware fill depend on bit 2 (i2s_fifo_enabled), bit 0 (i2s_drive), or seek-complete alone?
Only then decide whether the cdrom.c bit-2 gate is itself wrong. Do NOT trust the cdrom.c comments claiming "the GPU ISR re-enables I2CNTRL bit 2" as established fact -- the trace shows no such write occurs today, but whether it should occur is exactly candidate 1.

Mechanism B -- unpopulated 68K exception vectors in CD-BIOS mode

Affects: Battle Morph (bios) [class 2, pc_escape $8FBFB758] and Baldies (bios) [class 4, reclassified from video_stall/HARNESS_HANG]. CONFIRMED escape mechanism; exact exception trigger UNCONFIRMED.

Both titles pc-escape at exactly frame 437 from the same prev_pc, with no CD activity whatsoever beforehand:

Battle Morph [PC-OOB] first oob at frame 437 PC=$8CA88044   [OOB-FROZEN] prev_pc=$8020A2 -> oob_pc=$8CA88044
Baldies      [PC-OOB] first oob at frame 437 PC=$A7E08FA0   [OOB-FROZEN] prev_pc=$8020A2 -> oob_pc=$A7E08FA0
both         [CD-DIAG] ... seeks=0 ... dsaIRQs=0 ... cdPlaying=0 i2sEn=0

seeks=0, dsaIRQs=0, cdPlaying=0: the crash happens before any seek or CD read. This rejects the wrong-LBA / $2C00-TOC / 149-sector pregap-MSF hypothesis for these titles -- the MSF->LBA path never runs.
The harness [OOB-PREVBYTES $8020A2] label is misleading: with prev_pc >= 0x200000 the dump offset falls back to 0 (test/test_cd_bios_boot.c:231), so those 32 bytes are actually main RAM $000000-$00001F -- the 68K exception vector table: 00 00 00 00 | 00 E0 00 08 | C6 E5 74 7A | 65 2A 09 AF | A7 E0 8F A0 ... = SSP=$0, reset=$00E00008 (valid), then bus-error/address-error/ illegal-instruction vectors = PRNG garbage. Vector $10 (illegal instruction) = $A7E08FA0, which is exactly Baldies' escape target.
Confirmed in code: in CD-BIOS mode vjs.useJaguarBIOS = true (libretro.c:1070, bootConfig.showBootROM). src/core/jaguar.c:831 then copies only the first 8 bytes (SSP + reset PC) from the boot ROM into RAM, and the HLE exception-vector-stub block (src/core/jaguar.c:857-881, which fills vectors 2-255 with RTE stubs) is gated on !vjs.useJaguarBIOS and therefore skipped. The auth-bypass + boot-stub-injection boot flow never runs the real BIOS code that would populate them either. Result: any exception/interrupt after boot vectors through PRNG garbage and the 68K pc-escapes.

Baldies' target $A7E08FA0 is a direct hit on the dumped illegal-instruction vector (RAM offset $10) -- the smoking gun. Battle Morph's $8CA88044 does NOT appear in the 32-byte RAM[0..31] dump, so it would sit in an un-dumped higher vector slot (vectors 8-255, RAM $20-$3FF); Battle Morph is therefore confirmed by the shared code gate + identical crash signature, not by a located slot. Both titles inject large boot stubs (Battle Morph $65290 @ $004400, Baldies $4000 @ $004000) and take their fatal exception at exactly frame 437 (an identical deterministic onset across two unrelated games -- consistent with a common BIOS/timing-driven trigger rather than game data, though the trigger itself is not isolated here). An unexplained detail worth flagging: both games escape from the same prev_pc under the same deterministic PRNG-seeded vector table, yet land on different targets ($8CA88044 vs $A7E08FA0) -- i.e. they take different vector slots (or RAM at those slots had already been overwritten differently by the two boot stubs), which the 32-byte dump cannot distinguish.

Task 6B update — premise OVERTURNED by instruction-level disasm; this is a MASK, real cause is the $2C00 TOC

The Task-6 REQUIREMENT (disasm before trusting RTE stubs) was carried out. It rejects the "populate vectors" fix for these two titles. Essentials:

prev_pc=$8020A2 is a red herring: the harness samples PC once per frame (test_cd_bios_boot.c:198,255), so it only means "the 68K was in this BIOS ROM→RAM copy loop (cmpa.l a2,a0 / bcs.b, cannot fault) at the end of frame 436" — it is not instruction-adjacent to the fault.

The actual first fault (logged from cpuextra.c::Exception, since reverted) is a synchronous CPU trap the loaded boot stub executes on its own error path, not an async interrupt:

Baldies:      Exception nr=4  currpc=$004D54 op=$4AFC  (68000 ILLEGAL)   -> vec4 =$A7E08FA0 garbage
Battle Morph: Exception nr=10 currpc=$0507E2 op=$AFDE  (Line-A word)     -> vec10=$8CA71E78 garbage

Baldies: coherent stub code moveq #1,d0; jsr $4E18; tst.l d0; bpl $004D56; move.w #$C000,$F00058; ILLEGAL. $4E18 is a TOC scanner (movea.l #$2C08,a0; movem.l (a0)+,d2-d3; tst.l d2; beq → moveq #-1,d0). Our injected TOC (src/cd/jagcd_bios.c:96-121) puts the session marker as an 8-byte slot with first longword = 0 at $2C08, so the scanner's first tst.l d2 sees 0 and returns error (-1) → the stub deliberately traps via ILLEGAL. Root cause = the $2C00 TOC format is incompatible with the boot stub's scanner (see project_cd_toc_format.md), NOT the vectors.
Battle Morph: currpc=$0507E2 is inside a data region (7838 7834 7838 7834 … run) — PC has derailed into data and executes it until hitting the Line-A word. Uncharacterized derailment (PC-in-data proven; the derailing jump not traced), likely downstream of the same loaded-data/TOC path.
Anomaly resolved: different vectors (4 vs 10) and once-per-frame sampling landing mid-runaway. Vec 4 = $A7E08FA0 re-faults on itself ($8888 = illegal) so Baldies' sampled PC == the raw vector; Battle Morph's vec-10 target $8CA71E78 differs from its sampled $8CA88044 (which is absent from the entire pristine 2 MB PRNG image) because its runaway wandered before the frame-boundary sample.

Verdict: populating vectors MASKS. The existing HLE stub sets vector 4 = plain RTE; the 68000 stacks PC at the illegal instruction, so RTE returns to $004D54 and re-executes ILLEGAL forever — Baldies would turn pc_escape into a wedge while a "vectors populated" test goes GREEN (PR#170 pattern). The same holds for Battle Morph's Line-A trap (vector 10 also stacks the faulting PC on a 68000, so plain RTE re-executes $AFDE = wedge), and for the jagcd_hle.c:934-949 "skip-2" ILLEGAL-handler variant (ADDQ.L #2,(2,SP); RTE): skipping past the deliberate halt lets the stub continue with d0=-1 and garbage TOC state — an equally effective mask. The real fix is the $2C00 TOC layout (a different mechanism, needs the authoritative MiSTer format) — tracked separately, not under 6B. Task 6B outcome: BLOCKED with artifacts, no src/ change.

TOC defect spec for the fix task (what the disasm establishes; resolve the open points against MiSTer/BIOS-DSP ground truth, not guesswork):

Defect (1) — marker slot terminates the scan. The injector writes the session-2 marker as an 8-byte slot whose first longword is 0 at $2C08 (marker byte 01 at +4); Baldies' $4E18 scanner reads entries from $2C08 and treats a zero first longword as end-of-table → instant d0=-1.
Defect (2) — no entry carries the byte[4] key. Even past the marker, the scanner matches via rol.l #8,d3; cmp.b d0,d3 — i.e. it compares the search key against entry byte[4] (first byte of the second longword). jagcd_bios.c writes only bytes [0..3] = [track][min][sec][frm] and leaves [4..7] zero, so the key match can never succeed for any entry. A fix that only un-zeroes the marker still returns d0=-1.
Format-model question to resolve: per-entry session/type byte at offset [4] (structurally closer to what $4E18 expects — it counts occurrences of the key byte, consistent with "Nth entry of session S") vs the current standalone-marker-slot design. Decide from MiSTer ground truth, not by making one scanner happy.
Two-scanner requirement: any TOC layout must satisfy BOTH Baldies' $4E18 occurrence-counting scan AND Primal Rage's $0803E2 marker-then-NEXT-entry scan (project_cd_toc_format.md). The found-path payload — which MSF fields the stub consumes from the matched entry ($4E48: ror.l #8,d3; bsr $4E54 onward) — is still unknown; MiSTer/BIOS-DSP ground truth required before writing bytes.
Frame-437 determinism, crisply: the shared CD BIOS reaches the $050176 boot-stub injection hook at ~frame 437 in both titles; each injected stub then faults within its first frame of execution. The constant is a property of the deterministic BIOS boot sequence, not a timer/interrupt.

HLE-mode wrong-LBA spot check -- Iron Soldier 2 (hle)

Wrong-LBA hypothesis REJECTED; failure is a repeated-read continuation loop. IS2 hle scores 0/1 with not_thrash=0. The HLE_READ sequence:

CD_read: D1='D8A' MSF=50:00:01 LBA=224851 dest=$00CB00 size=$14FE  -> sentinel match at LBA 224857
CD_read: (identical) LBA=224851 ... -> repeated read -- resuming from LBA 224860 ... sentinel NOT found
CD_read: (identical) LBA=224851 ... -> repeated read -- resuming from LBA 224863 ... sentinel NOT found
... (+3 LBA per identical call, indefinitely: 224866, 224869, 224872, ...)

MSF 50:00:01 -> LBA ((50*60+0)*75 + 1) - 150 = 224851. Disc layout ([CD-LAYOUT]) puts track 12 (session 2 boot/data) at startLBA 224391, len 764 (ends 225155); 224851 is validly inside track 12 -- the LBA and MSF->LBA math are correct.
The game re-issues a byte-identical CD_read (same MSF/LBA/dest/size), but the jagcd_hle.c "repeated read" heuristic advances the resume LBA by +3 each call and reports "sentinel NOT found", so the game never receives the data it expects and retries forever. The bug is the HLE repeated-read/continuation logic, not MSF->LBA mapping.

Task 6C fix -- delete the continuation heuristic (CD_read is idempotent)

Contract. CD_read ($303C) is a discrete, self-contained read: D0 (packed MSF) fully specifies the source position, A0/A1 the destination range, D1 the sync marker to lock onto in the I2S stream after the seek. A re-issued byte-identical call re-seeks to the SAME position and reproduces the SAME data. There is no per-call "continuation" pointer. This holds under both readings of the real service (stateless-idempotent, or a BIOS that streams internally): a repeat with an identical MSF is a restart from the same position either way. The only reading in which the removed heuristic is correct -- a BIOS that ignores the passed MSF and advances an internal pointer -- is self-contradictory (why pass an MSF at all) and unsupported.

Where the heuristic came from. It was a tuned guess, not a disassembled contract: the writeback hle_next_lba = scanLBA + s (post-transfer LBA) plus the hle_have_last "repeated read -- resuming from LBA" block, extended by commit 4844f6b into the raw-fallback and counter-ID paths. The +3/call IS2 saw is just s = ceil(0x14FE / 2352) = 3 sectors per transfer. The comment admitted it "mimics the I2S stream that real hardware would still be feeding" -- a fabricated streaming model layered on a stateless service.

LBA trace, IS2 (hle), before vs after (VJ_CD_TRACE/HLE_LOG, identical D0=$00320001 D1=$4438410B dest=$00CB00 size=$14FE every call):

BEFORE (heuristic):                          AFTER (idempotent):
call 1: sync block @ LBA 224857 off 278  ->  call 1: sync block @ LBA 224857 off 278
call 2: resume 224860, sentinel NOT found    call 2: sync block @ LBA 224857 off 278
call 3: resume 224863, sentinel NOT found    call 3: sync block @ LBA 224857 off 278
call N: resume 224857+3N, raw garbage        call N: sync block @ LBA 224857 off 278
   -> dest $00CB00 corrupted on every repeat     -> dest $00CB00 stable (correct) every repeat

RED/GREEN. test/test_cd_hle_idempotent.c runs IS2 (hle) and counts how many times the destination region $00CB00..$00DFFF changes content after it is first populated. RED (heuristic present): 27 changes over 300 frames. GREEN (heuristic removed): 0 changes. Golden-free, deterministic, game-agnostic.

Matrix (current HEAD, before vs after the fix, matched frame budgets -- the older rows in the table below predate the HEAD~1 TOC-writer rewrite and are stale for Baldies hle):

Row	Mode	Frames	BEFORE (heuristic)	AFTER (fix)
Iron Soldier 2	hle	3000	FAIL not_thrash=0 unique=3 $007416	FAIL not_thrash=0 unique=3 $007416
Baldies	hle	3000	FAIL unique=3 $05FE0C	FAIL unique=3 $05FE0C
Primal Rage	hle	3000	PASS unique=14 $00419E	PASS unique=14 $00419E
BrainDead 13	hle	3000	PASS unique=6 $1243BC	PASS unique=6 $1243BC
Dragon's Lair	hle	3000	PASS $005412	PASS unique=60 $005412
Primal Rage	bios	300	PASS/BIOS_INTRO $0059B0	PASS/BIOS_INTRO $0059B0
Baldies	bios	300	PASS/BIOS_INTRO $0059B0	PASS/BIOS_INTRO $0059B0

All controls unchanged. bios rows are structurally unaffected -- the heuristic lives only on the HLE path (cd_boot_strategy_hle).

Honest outcome (necessary, not sufficient). The fix provably stops the corruption -- IS2's first read stays correct and idempotent on every repeat -- but IS2 (hle) does NOT boot past the loop: its harness verdict is unchanged (not_thrash=0, 3 game-code PCs $007416/$006F78/$006B74, final_pc=$007416). CD_poll reports completion (A0 reaches the GPU data pointer), so the residual loop is downstream game logic re-issuing the read -- a separate defect from the continuation heuristic (candidate: the game validates loaded state / awaits a subsequent read that isn't modelled). Not re-tuned to force a boot. Baldies (hle) does not issue repeated-identical reads at HEAD, so it neither regressed nor advanced -- the "Baldies shares the heuristic" hypothesis is not supported by the trace.

Summary

Class	Title(s)	Verdict	Mechanism
1	Primal Rage (bios)	deadlock FIXED (Task 6A) -- attribution was our IRQ0-vs-IRQ1 routing	Mechanism A resolved: BUTCH interrupt now lands on GPU IRQ1 ($F03010) where the BIOS CD-ISR lives; FIFO fills/drains live. Next blocker: boot-stub sentinel scan (separate mechanism)
3	Highlander (bios)	deadlock FIXED (Task 6A)	Same fix as class 1; fifo_drains now non-zero (was 0). Downstream progress blocked same as class 1
3	Iron Soldier 2 (bios)	deadlock FIXED (Task 6A) / drain-chain caveat moot	Same fix; the previously-unresolved drain/refill oddity was a symptom of the never-dispatched ISR
2	Battle Morph (bios)	CONFIRMED escape / UNCONFIRMED trigger	Unpopulated 68K exception vectors in CD-BIOS mode -> exception vectors through garbage (Mechanism B); pre-CD-read, wrong-LBA rejected
4	Baldies (bios)	CONFIRMED	Same as class 2 (Mechanism B); reclassified from video_stall/HANG. video_stall is downstream of the frame-437 escape
--	Iron Soldier 2 (hle)	wrong-LBA REJECTED	LBAs correct; repeated-read continuation loop in jagcd_hle.c

Re-run notes -- 2026-07-14, Task 8 fix (GPU delay-slot IRQ dispatch clobber)

Seven rows deleted + re-run after fix(gpu): don't clobber IRQ dispatch raised in a branch delay slot (3000 frames, same knobs). Root cause and byte-level evidence in docs/cd-diagnosis/braindead13-second-wedge-diagnosis.md: the CD streaming ISR epilogue (JUMP T,(Rret) with an IMASK-clearing G_FLAGS store in the delay slot) had its interrupt dispatch overwritten by gpu_opcode_jump/jr applying the branch target after the delay slot ran -- IMASK stayed set forever and the CD ISR never fired again. This was the post-first-load wedge: cd_seek_wedge / video_stall with G_FLAGS.IMASK stuck and the GPU spinning in a foreground mailbox loop ($F03168 Primal Rage 2nd engine, $F0327x BrainDead / Dragon's Lair / Space Ace class). Contract test: test/test_cd_second_transfer (VJ_FIFO_DISC-gated).

BrainDead 13 (bios) -- ADVANCED. Was: video_stall frame=739, final_pc=$00361E, dsp_run=0 (the user-confirmed on-device logo wedge). Now: logo load completes, the game issues a SECOND ~1 MB FMV streaming read and completes it (38,915 fifo drains, IMASK clear), dsp now running, final_pc=$00B2FA (new game-code band), unique_pcs 23->43, payload 25 kB. The remaining video_stall is the game's frame-kick wait (headless-fb class).
Primal Rage (bios) -- ADVANCED to the HLE oracle's state. Was: cd_seek_wedge, drains frozen 12,857 mid-scan, gpu wedged $F03168. Now: transfer #2 runs to completion (drains 17,996; the engine's ISR signals done by clearing BUTCH bit 0), gpu waits at $F031F6 with IMASK clear -- byte-for-byte the same wait state as the Primal Rage (hle) control row, payload 6790B vs hle 6679B. The row logged cd_seek_wedge at the time: drains legitimately stop after transfer completion (BUTCH bit 0 off), which the watchdog could not then distinguish from a freeze. Fixed in #741 -- the watchdog now requires BUTCH's master + FIFO interrupts to still be armed and I2CNTRL FIFO data enabled.
Highlander (bios) -- ADVANCED. Was: cd_seek_wedge with drains frozen at 54 right after the first fills (same clobber, first ISR invocations). Now: no seek wedge, gpu active at $F031C2, dsp running, final_pc=$0036BE (CD service band, PC-SET suppressed at unique_pcs=38).
Dragon's Lair (bios) (control, but itself an IMASK-wedge-class title: its old stall had gpu_pc=$F0327A) -- moved FORWARD: BIOS_INTRO $0060D8 -> CD service band $0036B8, unique_pcs 38->51, payload 23.4->27.2 kB, stall deferred 840->1008.
Iron Soldier 2 (bios) -- byte-identical row (its blocker is a different, earlier mechanism; unaffected as expected).
Controls Primal Rage (hle), Iron Soldier 2 (hle) -- byte-identical.

| Baldies (USA) (Rev 1).cue | bios | 1/1 | GAME_CODE | [CRASH-DETECT] video_stall frame=865 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B086 dsp_run=1 | [PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=11957B unique_pcs=38 final_pc=$060106 | | BrainDead 13 (USA).cue | bios | 0/1 | GAME_CODE | [CRASH-DETECT] video_stall frame=957 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B084 dsp_run=1 | [FAIL] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=24918B unique_pcs=43 final_pc=$00B2FA | | Dragon's Lair (USA).cue | bios | 1/1 | ? (BIOS service band $0-$3FFF; PC-SET suppressed, unique_pcs>32 -- game handoff undetermined) | [CRASH-DETECT] video_stall frame=1008 fb_hash=$4DE01878 unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B082 dsp_run=1 | [PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=27155B unique_pcs=51 final_pc=$0036B8 | | Highlander - The Last of the MacLeods (USA).cue | bios | 1/1 | ? (BIOS service band $0-$3FFF; PC-SET suppressed, unique_pcs>32 -- game handoff undetermined) | [CRASH-DETECT] video_stall frame=920 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F031C2 gpu_run=1 dsp_pc=$00F1B07E dsp_run=1 | [PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=2726B unique_pcs=38 final_pc=$0036BE | | Iron Soldier 2 (USA) (Songbird).cue | hle | 0/1 | GAME_CODE | (none) | [FAIL] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=0 ram_payload=474B unique_pcs=3 final_pc=$007416 | | Iron Soldier 2 (USA) (Songbird).cue | bios | 1/1 | ? (BIOS service band $0-$3FFF; PC-SET showed no game-band entries) | [CRASH-DETECT] video_stall frame=760 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B04A dsp_run=1 | [PASS] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=584B unique_pcs=31 final_pc=$0036B8 | | Primal Rage (USA).cue | hle | 1/1 | GAME_CODE | [CRASH-DETECT] video_stall frame=530 fb_hash=$BFFF6779 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0AE dsp_run=1 | [PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=6679B unique_pcs=14 final_pc=$00419E | | Primal Rage (USA).cue | bios | 1/1 | ? (BIOS service band $0-$3FFF; PC-SET suppressed, unique_pcs>32 -- game handoff undetermined) | [CRASH-DETECT] cd_seek_wedge frame=958 seek_starts=2 seek_dones=2 fifo_drains=17996 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0B0 dsp_run=1 | [PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=6790B unique_pcs=45 final_pc=$0036B8 |

Re-run notes -- 2026-07-15, lost-wakeup fix (event-scheduled GPU CPUINT)

Full-table re-run (all 21 rows deleted + regenerated, 3000 frames, 120 s cap) after the 68K/GPU lost-wakeup fix: GPU-raised CPUINT (G_CTRL bit 1, who==GPU, 68K not already stopped) is now delivered through the event scheduler at (slice budget + GPU cycles consumed) usec instead of synchronously, so the one-shot wakeup can no longer be consumed before the 68K executes the stop of a command/stop/wait handshake. Raw logs: /tmp/cdmx_postfix.

BrainDead 13 (bios) -- THE TARGET -- ADVANCED. Was: score 0/1, final_pc=$00B2FA (the documented lost-wakeup deadlock loop), not_loop=0, video_stall frame=957. Now: score 1/1, not_loop=1, unique_pcs 43->59, ram_payload 24918->25309B, no video_stall; final_pc=$0036B8 is the BIOS CD-read service band, i.e. the game is alive and streaming. A PC-band sampler (3000 frames) confirms game-band execution from frame 438 and test/test_cd_lost_wakeup (new, in make test) passes with a worst halted-no-IRQ window of 1 frame (was: permanent).
Stage-column caveat for the three GAME_CODE -> ? moves (Baldies bios, BrainDead 13 bios, and BIOS_INTRO -> ? for Hover Strike / Space Ace bios): the stage classifies the FINAL-frame PC only. Post-fix these titles spend most frame boundaries inside the BIOS CD-read service band ($0-$3FFF) because more interrupts flow and more disc I/O happens. Direct PC-band sampling over all 3000 frames shows Baldies (bios) in game-band code for 522/3000 samples (first at frame 566) -- GAME_CODE is still reached; the ? is a sampling artifact, not a regression. Counter evidence agrees: Baldies fifoIRQs 45895 with fifoReads 14.7M; Space Ace (bios) advanced from seeks=1 to seeks=7 / drains=173466.
No row moved backward on score or on evidence-verified stage; baldies.cdi LOAD_FAIL and Battle Morph (bios) pc_escape are unchanged pre-existing backlog items (see handoff doc section 5).

Re-run notes -- 2026-07-15 (2), GPU internal-RAM LOADB/LOADW fix

Full-table re-run after fix(gpu) f13892b (internal byte/word loads perform a 32-bit read). All 21 rows IDENTICAL in score and stage to the previous run -- expected: the fix corrects decoded pixel VALUES (Cinepak FMV), which the boot-stage taxonomy cannot observe. Visual confirmation is in docs/cd-diagnosis/fmv-corruption-diagnosis.md (BrainDead 13 / Dragon's Lair FMV now render pixel-perfect via cd_visual_verify screenshots). Raw logs: /tmp/cdmx_loadb.

Re-run notes -- 2026-07-15, I2CNTRL status-bit fix (`f74ce09`)

Full 21-row re-run after masking the read-only FIFO-not-empty status bit (b4) on I2CNTRL writes. The pre-fix baseline had six bios rows wedged at final_pc $36B2-$36C2 -- all the SAME BIOS end-of-transfer FIFO flush loop ("read FIFO_DATA; btst #4,I2CNTRL; bne", BIOS service code resident at $3600) spinning on the stuck status bit. Post-fix:

Baldies bios: "? BIOS service band" + cd_seek_wedge -> 1/1 GAME_CODE, no watchdog
Highlander bios: "? BIOS service band" + video_stall -> 1/1 GAME_CODE
Iron Soldier 2 bios: "? BIOS service band" -> 1/1 GAME_CODE
BrainDead 13 / Primal Rage bios: "? BIOS service band" -> BIOS_INTRO (final-frame sampling label shift within the known artifact class; scores unchanged 1/1)
Dragon's Lair bios: still "?" at the 3000-frame horizon, but now streams five further transfer cycles (seeks 2 -> 10) and reaches its in-game CD error handler (~frame 2760) -- next diagnosis target
Everything else: unchanged

Re-run notes -- 2026-07-15, FIFO transfer-rate fix (real 2x drive speed)

Full 21-row re-run after pacing FIFO refills at the hardware rate (352,800 B/s; error-diffused 2.85-tick period) instead of the fixed 5-tick ~201 KB/s. The ReadySoft schedule-watchdog class (Dragon's Lair / Space Ace "error reading CD" + looping-FMV retry) is fixed -- both titles stream FMV continuously through 4200-frame visual runs with no watchdogs. Matrix deltas vs the I2CNTRL-fix run:

Hover Strike bios: "? BIOS service band" -> 1/1 GAME_CODE
Space Ace bios: "? BIOS service band" -> 1/1 GAME_CODE
Dragon's Lair bios: "?" -> BIOS_INTRO (final-frame sampling label; visual run confirms continuous in-game FMV)
Space Ace hle recorded HARNESS_HANG under heavy machine load (concurrent visual runs); direct re-run exits normally with the known pre-existing HLE stall (video_stall + silent audio, same as Dragon's Lair hle)
Everything else unchanged

Re-run notes -- 2026-07-16, SCLK synchronous-SSI-IRQ fix (`b63fb7d`)

Full 21-row re-run (fresh table, 3000 frames, 120 s cap) at b63fb7d (fix(jerry): SCLK write no longer fires the SSI interrupt synchronously): writing SCLK ($F1A150) no longer invokes JERRYI2SCallback() inside the programming store; the first SSI interrupt is deferred one full I2S frame period (JERRYRescheduleI2S), so a DSP audio module re-uploaded mid-boot (park DSP -> upload -> D_PC -> DSPGO) finishes its register init before its I2S handler can fire with the OLD module's registers still live. That stale dispatch was the mechanism behind the recurring dsp_pc_escape $FFFFCE03 / video_stall class (the handler mixed a CD sample through stale r13 over its own code at $F1B07C). Note this table also folds in the five CD fixes landed since the previous (FIFO transfer-rate) table was generated (c336839, 0f22fac, 746cf4a, 0e30eed, 874330f) -- their combined matrix-visible delta is nil outside the rows called out below.

Highlander (hle) -- THE TARGET -- dsp_pc_escape GONE. Was: [CRASH-DETECT] dsp_pc_escape frame=128 pc=$FFFFCE03, unique_pcs=5, final_pc=$008D66. Now: no crash-detect signature at all, unique_pcs=10 (PC-SET all game-band: $00412C $005876 $008BBC ...), final_pc=$008BBC, seeks=7, i2sEn=1, fifoIRQs=178636 -- the DSP survives the module re-upload and the CD stream keeps flowing.
Space Ace (hle) -- the previous table's HARNESS_HANG (recorded under heavy machine load, already annotated there as a flake) recovered to the known 1/1 GAME_CODE state (video_stall f529, final_pc=$0055CC, unique_pcs=68), matching that section's direct re-run.
All other 19 rows byte-identical to the previous table, including the two rows this re-run targeted for confirmation: Dragon's Lair (bios) (BIOS_INTRO, no watchdog, final_pc=$004C10) and Highlander (bios) (GAME_CODE, final_pc=$0088D8).
Visual/audio confirmation (cd_visual_verify, 3000 frames, bios mode, this exact build) -- the boot-stage taxonomy cannot see "renders with audio", so recorded here: Dragon's Lair streams FMV continuously through frame 2999 (motion in every 60-frame window, avg change up to 25%, 79.5% non-black), audio RMS 1234, zero crash-detect signatures. Highlander renders ~98% non-black with sustained motion, audio RMS 1526.
Remaining open bug -- Highlander (bios) cd_seek_wedge. The signature fires at frame 1743 (seek_starts=4 seek_dones=4, fifo_drains frozen at 92617) and AGAIN at frame 2704 (seeks 6/6, drains 149967); the visual run shows rendering, motion, and audio continue across both, and the run ends with seeks=13, cdPlaying=1, i2sEn=1. Disc activity resumes after each signature, so this is an intermittent drain stall (or a watchdog false-positive during paused-drive gaps, same class as the Primal Rage post-transfer note in the Task 8 section) rather than a terminal wedge -- next diagnosis target.

Title	Mode	Score	Stage	Watchdog	PC evidence
Baldies (USA) (Rev 1).cue	hle	0/1	GAME_CODE	[CRASH-DETECT] video_stall frame=302 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B088 dsp_run=1	[FAIL] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=0 not_thrash=0 ram_payload=11731B unique_pcs=3 final_pc=$05FE0C
Baldies (USA) (Rev 1).cue	bios	1/1	GAME_CODE	[CRASH-DETECT] cd_seek_wedge frame=990 seek_starts=2 seek_dones=2 fifo_drains=45895 unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B086 dsp_run=1	[PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=21698B unique_pcs=58 final_pc=$05FE8A
Battle Morph (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=1359 fb_hash=$114AB567 unchanged for 300 frames gpu_pc=$00F031D8 gpu_run=0 dsp_pc=$00F1B57E dsp_run=1	[PASS] Battle Morph (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=22157B unique_pcs=126 final_pc=$0046DC
Battle Morph (USA).cue	bios	0/1	? (pc_escape)	(none)	[FAIL] Battle Morph (USA).cue : pc_in_ram=0 not_loop=1 not_thrash=1 ram_payload=32704B unique_pcs=256+ final_pc=$8FBFB758
BrainDead 13 (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=1531B unique_pcs=6 final_pc=$1243BC; PC-SET game-band: $12438A $1243BC $12438C $12562C $1243BA
BrainDead 13 (USA).cue	bios	1/1	BIOS_INTRO	[CRASH-DETECT] cd_seek_wedge frame=1107 seek_starts=2 seek_dones=2 fifo_drains=38915 unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=27663B unique_pcs=65 final_pc=$004FD0
Dragon's Lair (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=412 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14268B unique_pcs=60 final_pc=$005412
Dragon's Lair (USA).cue	bios	1/1	BIOS_INTRO	(none)	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29338B unique_pcs=65 final_pc=$004C10
Highlander - The Last of the MacLeods (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14691B unique_pcs=10 final_pc=$008BBC
Highlander - The Last of the MacLeods (USA).cue	bios	1/1	GAME_CODE	[CRASH-DETECT] cd_seek_wedge frame=1743 seek_starts=4 seek_dones=4 fifo_drains=92617 unchanged for 300 frames gpu_pc=$00F031C2 gpu_run=1 dsp_pc=$00F1B07C dsp_run=1	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=17876B unique_pcs=62 final_pc=$0088D8
Hover Strike - Unconquered Lands (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=15997B unique_pcs=19 final_pc=$065B36
Hover Strike - Unconquered Lands (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=17157B unique_pcs=71 final_pc=$05DA9E
Iron Soldier 2 (USA) (Songbird).cue	hle	0/1	GAME_CODE	(none)	[FAIL] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=0 ram_payload=474B unique_pcs=3 final_pc=$007416
Iron Soldier 2 (USA) (Songbird).cue	bios	1/1	GAME_CODE	(none)	[PASS] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29870B unique_pcs=39 final_pc=$00B542
Primal Rage (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=530 fb_hash=$BFFF6779 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0AE dsp_run=1	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=6679B unique_pcs=14 final_pc=$00419E
Primal Rage (USA).cue	bios	1/1	BIOS_INTRO	[CRASH-DETECT] cd_seek_wedge frame=884 seek_starts=2 seek_dones=2 fifo_drains=17996 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0AE dsp_run=1	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=25894B unique_pcs=78 final_pc=$0044D0
Space Ace (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=529 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F0327A gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14565B unique_pcs=68 final_pc=$0055CC
Space Ace (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29417B unique_pcs=75 final_pc=$00B87A
baldies.cdi	hle	0/1	LOAD_FAIL (harness crash)	(none)	[CRASH] baldies.cdi : child died with signal 11 (Segmentation fault: 11)
baldies.cdi	bios	0/1	LOAD_FAIL (harness crash)	(none)	[CRASH] baldies.cdi : child died with signal 11 (Segmentation fault: 11)

Raw per-run logs are not committed; re-run the script to regenerate.

Re-run notes -- 2026-07-16, SCLK sync-SSI-IRQ fix + SSI head split + drive-silence gate

Regenerated @ ca82c5b (fixes since the 07-15 table: b63fb7d SCLK no longer fires the SSI interrupt synchronously; 0e30eed independent SSI read head; 874330f paused/seeking drive delivers I2S silence; 746cf4a TOC track durations; ca82c5b matrix tooling -- bare make test/<bin> linked an unexecutable dylib via the built-in rule, which is why the first 07-16 regeneration attempt produced all-? rows).

9 of 21 rows changed; no regressions:

Highlander hle: dsp_pc_escape $FFFFCE03 watchdog -> (none) -- the SCLK synchronous-IRQ fix, exactly as verified in b63fb7d.
Dragon's Lair bios: video_stall f1008 -> (none), stage now BIOS_INTRO (was ? service band).
Hover Strike bios / Iron Soldier 2 bios / Space Ace bios: watchdog signatures cleared, stage now GAME_CODE (handoff determined -- these previously sat in ? (BIOS service band)).
Highlander bios: video_stall f920 -> cd_seek_wedge f1743 -- runs much further; the remaining wedge is the known open item noted in b63fb7d's verification.
Baldies bios: stage ? -> GAME_CODE; cd_seek_wedge remains.
BrainDead 13 bios / Primal Rage bios: stage ? -> BIOS_INTRO reclassification; cd_seek_wedge remains at slightly different frames.

Several PC-SET suppressed indeterminates resolved to real stages this run -- fewer early wedges means the harness now emits the PC-SET dump.

Re-run notes -- 2026-07-16, streamed HLE CD_read (`a1f1ac1`)

HLE CD_read now streams at the double-speed drive rate (352,800 B/s) instead of completing instantly. Instant delivery overwrote game code that was still executing inside the destination window (Hover Strike's LVL6 mission load: 2/2 scripted-input lockups before, 0/16 after -- the "lockup when skipping cutscenes / starting a mission" reports).

Deltas vs the previous table; no regressions:

Baldies hle: 0/1 FAIL (not_thrash=0, unique_pcs=3) -> 1/1 GAME_CODE (unique_pcs=19) -- the game's load sequencing also relied on observing streaming progress.
Battle Morph hle / Primal Rage hle: PASS as before, richer PC sets (126->147 / 14->21 unique).
Everything else identical.

Title	Mode	Score	Stage	Watchdog	PC evidence
Baldies (USA) (Rev 1).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=369 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B088 dsp_run=1	[PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=11739B unique_pcs=19 final_pc=$05FE0C
Baldies (USA) (Rev 1).cue	bios	1/1	GAME_CODE	[CRASH-DETECT] cd_seek_wedge frame=990 seek_starts=2 seek_dones=2 fifo_drains=45895 unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B086 dsp_run=1	[PASS] Baldies (USA) (Rev 1).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=21698B unique_pcs=58 final_pc=$05FE8A
Battle Morph (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=302 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F030C8 gpu_run=1 dsp_pc=$00F1C072 dsp_run=1	[PASS] Battle Morph (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=22157B unique_pcs=147 final_pc=$0046DC
Battle Morph (USA).cue	bios	0/1	? (pc_escape)	(none)	[FAIL] Battle Morph (USA).cue : pc_in_ram=0 not_loop=1 not_thrash=1 ram_payload=32704B unique_pcs=256+ final_pc=$8FBFB758
BrainDead 13 (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=1531B unique_pcs=10 final_pc=$12438A; PC-SET game-band: $124342 $12438A $1243BC $12438C $1257E8 $124396 $12434C $1243BA $125902
BrainDead 13 (USA).cue	bios	1/1	BIOS_INTRO	[CRASH-DETECT] cd_seek_wedge frame=1107 seek_starts=2 seek_dones=2 fifo_drains=38915 unchanged for 300 frames gpu_pc=$00F03270 gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] BrainDead 13 (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=27663B unique_pcs=65 final_pc=$004FD0
Dragon's Lair (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=412 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F03276 gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14268B unique_pcs=60 final_pc=$005412
Dragon's Lair (USA).cue	bios	1/1	BIOS_INTRO	(none)	[PASS] Dragon's Lair (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29338B unique_pcs=65 final_pc=$004C10
Highlander - The Last of the MacLeods (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14691B unique_pcs=20 final_pc=$00826E
Highlander - The Last of the MacLeods (USA).cue	bios	1/1	GAME_CODE	[CRASH-DETECT] cd_seek_wedge frame=1743 seek_starts=4 seek_dones=4 fifo_drains=92617 unchanged for 300 frames gpu_pc=$00F031C2 gpu_run=1 dsp_pc=$00F1B07C dsp_run=1	[PASS] Highlander - The Last of the MacLeods (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=17876B unique_pcs=62 final_pc=$0088D8
Hover Strike - Unconquered Lands (USA).cue	hle	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=7254B unique_pcs=52 final_pc=$065B36
Hover Strike - Unconquered Lands (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Hover Strike - Unconquered Lands (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=17157B unique_pcs=71 final_pc=$05DA9E
Iron Soldier 2 (USA) (Songbird).cue	hle	0/1	GAME_CODE	(none)	[FAIL] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=0 ram_payload=474B unique_pcs=3 final_pc=$007416
Iron Soldier 2 (USA) (Songbird).cue	bios	1/1	GAME_CODE	(none)	[PASS] Iron Soldier 2 (USA) (Songbird).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29870B unique_pcs=39 final_pc=$00B542
Primal Rage (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=620 fb_hash=$BFFF6779 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0AC dsp_run=1	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=0 not_thrash=1 ram_payload=6679B unique_pcs=21 final_pc=$00419E
Primal Rage (USA).cue	bios	1/1	BIOS_INTRO	[CRASH-DETECT] cd_seek_wedge frame=884 seek_starts=2 seek_dones=2 fifo_drains=17996 unchanged for 300 frames gpu_pc=$00F031F6 gpu_run=1 dsp_pc=$00F1B0AE dsp_run=1	[PASS] Primal Rage (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=25894B unique_pcs=78 final_pc=$0044D0
Space Ace (USA).cue	hle	1/1	GAME_CODE	[CRASH-DETECT] video_stall frame=529 fb_hash=$292A41FD unchanged for 300 frames gpu_pc=$00F0327A gpu_run=1 dsp_pc=$00F1B082 dsp_run=1	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=14565B unique_pcs=68 final_pc=$0055CC
Space Ace (USA).cue	bios	1/1	GAME_CODE	(none)	[PASS] Space Ace (USA).cue : pc_in_ram=1 not_loop=1 not_thrash=1 ram_payload=29417B unique_pcs=75 final_pc=$00B87A
baldies.cdi	hle	0/1	LOAD_FAIL (harness crash)	(none)	[CRASH] baldies.cdi : child died with signal 11 (Segmentation fault: 11)
baldies.cdi	bios	0/1	LOAD_FAIL (harness crash)	(none)	[CRASH] baldies.cdi : child died with signal 11 (Segmentation fault: 11)

Raw per-run logs are not committed; re-run the script to regenerate.

Raw per-run logs: /Users/jmattiello/vj-remote/validation/release-3.10.0-final-b5e40090/cd (not committed; re-run to regenerate).

90 KiB Raw Blame History

CD boot matrix

How to re-run

Legend

Baseline notes -- 2026-07-13, branch feature/jaguar-cd-support @ cbd2518

Results

Regenerated on ab77d16 -- 2026-10-07 (v3.7.1 release gate)

Regenerated on 2f913d8 -- 2026-10-02 (#735)

Myst CHD joined the sweep -- 2026-10-01

Myst joined the sweep -- 2026-07-30

baldies.cdi must be regenerated with cue2cdi, not taken from the ROM set

Re-run notes -- 2026-07-29, branch fix/cd-fmv-bios-handoff @ 17375c3

Why these three rows were never stuck

Re-run notes -- 2026-07-27, branch feature/jaguar-cd-support @ 73438ba

Re-run notes -- 2026-07-13, Task 6A fix (BUTCH -> GPU IRQ1 routing)

Re-run notes -- 2026-07-14, Task 6D fix ($2C00 TOC track-indexed layout)

Diagnosis (Task 5) -- 2026-07-13, branch feature/jaguar-cd-support @ 5827940

Mechanism A -- I2S FIFO never enabled after seek (FIFO-wait deadlock)

Mechanism B -- unpopulated 68K exception vectors in CD-BIOS mode

Task 6B update — premise OVERTURNED by instruction-level disasm; this is a MASK, real cause is the $2C00 TOC

HLE-mode wrong-LBA spot check -- Iron Soldier 2 (hle)

Task 6C fix -- delete the continuation heuristic (CD_read is idempotent)

Summary

Re-run notes -- 2026-07-14, Task 8 fix (GPU delay-slot IRQ dispatch clobber)

Re-run notes -- 2026-07-15, lost-wakeup fix (event-scheduled GPU CPUINT)

Re-run notes -- 2026-07-15 (2), GPU internal-RAM LOADB/LOADW fix

Re-run notes -- 2026-07-15, I2CNTRL status-bit fix (f74ce09)

Re-run notes -- 2026-07-15, FIFO transfer-rate fix (real 2x drive speed)

Re-run notes -- 2026-07-16, SCLK synchronous-SSI-IRQ fix (b63fb7d)

Re-run notes -- 2026-07-16, SCLK sync-SSI-IRQ fix + SSI head split + drive-silence gate

Re-run notes -- 2026-07-16, streamed HLE CD_read (a1f1ac1)

90 KiB

Raw Blame History

Baseline notes -- 2026-07-13, branch feature/jaguar-cd-support @ `cbd2518`

Regenerated on `ab77d16` -- 2026-10-07 (v3.7.1 release gate)

Regenerated on `2f913d8` -- 2026-10-02 (#735)

`baldies.cdi` must be regenerated with `cue2cdi`, not taken from the ROM set

Re-run notes -- 2026-07-29, branch fix/cd-fmv-bios-handoff @ `17375c3`

Re-run notes -- 2026-07-27, branch feature/jaguar-cd-support @ `73438ba`

Diagnosis (Task 5) -- 2026-07-13, branch feature/jaguar-cd-support @ `5827940`

Re-run notes -- 2026-07-15, I2CNTRL status-bit fix (`f74ce09`)

Re-run notes -- 2026-07-16, SCLK synchronous-SSI-IRQ fix (`b63fb7d`)

Re-run notes -- 2026-07-16, streamed HLE CD_read (`a1f1ac1`)