Joe MattielloandClaude Fable 5
9341fe3009
* test(cd): document --option passthrough + CD boot-mode gotcha in cd_visual_verify The harness already parses --option KEY=VALUE and serves it in GET_VARIABLE, but cd_visual_verify's usage text and header never mentioned it -- and --bios does NOT switch the CD boot path (that is the core option virtualjaguar_cd_boot_mode, default hle, which overrides the cart-BIOS setting for CD content). Every prior cd_visual_verify sweep therefore ran HLE regardless of --bios. Advertise --option in the usage/help text with the cd_boot_mode example, and echo option overrides at startup so saved logs record which boot mode a run actually requested. Co-Authored-By: Claude Fable 5 <[email protected]> * docs: FMV CD titles verified under real-BIOS boot mode (evidence table) Runs the previously-untested virtualjaguar_cd_boot_mode=bios path over the 5 FMV titles (3000f sweep + 15000f long probes of Dragon's Lair and Space Ace in both modes) with per-run boot-mode confirmation from the core log, motion timelines, audio RMS, and visually-inspected screenshots. All 5 pass; no late failures; all watchdog fires corroborated benign. Also baselines the exact failure signatures of the 4 damaged V2 CDI rips in both modes as before-evidence for the rip repair track. Co-Authored-By: Claude Fable 5 <[email protected]> --------- Co-authored-by: Claude Fable 5 <[email protected]>
8.5 KiB
FMV CD titles — real-BIOS boot mode verification (evidence)
Date: 2026-08-07. Core build: v3.1.0 be95df3 (develop). Harness:
test/tools/cd_visual_verify (headless; motion timeline + audio RMS +
periodic screenshots). Host: macOS arm64.
Why this sweep exists
Users report FMV titles broken in our core but fine in BigPEmu. A prior
3000-frame sweep passed all 5 titles — but it ran HLE boot mode every
time: cd_visual_verify --bios only sets virtualjaguar_bios (cart boot
ROM), while for CD content the boot path is chosen by the core option
virtualjaguar_cd_boot_mode (hle/auto/bios, default hle), which
overrides it. Real-BIOS mode was therefore never actually tested.
This sweep runs the real-BIOS path explicitly via
--option virtualjaguar_cd_boot_mode=bios and confirms per-run from the
core's own INFO log ([BOOT] CD game, mode=BIOS (BIOS image staged) +
[CD] Boot path: REAL BIOS at $802000) that the BIOS path was active.
All bios-mode runs used the embedded retail CD BIOS.
Verdict summary
Real-BIOS mode plays all 5 FMV titles correctly. Every title boots
through the CD BIOS animation, extracts its boot stub, reaches game code,
and shows sustained recognizable FMV imagery with audio. No late
failures in 15000-frame (≈4 min) probes of Dragon's Lair and Space Ace in
either mode. The 4 damaged V2 CDI rips fail identically in a
well-diagnosed way (missing boot header in the rip — see baseline below),
unchanged from the known docs/memory state.
Main sweep — 3000 frames, virtualjaguar_cd_boot_mode=bios
Motion% = frames with >0.5% pixels changed / measured frames, whole run (includes BIOS animation, loads, and static menu screens — FMV titles idle on static cards between clips, so 100% is not expected). RMS is the harness average over its recorded audio-frame window.
| Title | Mode | Verdict | Motion % | Peak nonblack | Audio RMS (nonsilent/total samples) | Watchdog lines | Evidence |
|---|---|---|---|---|---|---|---|
| Dragon's Lair | bios | PASS — FMV plays (Dirk + Singe scenes) | 64.1% | 79.5% | 1230 (1.72M/2.4M) | none | shots/dragons_lair_bios_3000/frame_01500.png, _02700.png |
| Space Ace | bios | PASS — FMV plays (Borf close-up) | 63.9% | 79.5% | 1753 (1.88M/2.4M) | none | shots/space_ace_bios_3000/frame_02100.png |
| BrainDead 13 | bios | PASS — reaches title/menu (castle-on-moon art), attract resumes | 20.5% | 78.5% | 2650 (2.05M/2.4M) | cd_seek_wedge f1107, video_stall f1529, cd_seek_wedge f2372 — all benign, see below |
shots/brain_dead_13_bios_3000/frame_01800.png, _02700.png |
| Blue Lightning | bios | PASS — FMV plays (aircraft over desert) | 27.7% | 94.9% | 1258 (2.12M/2.4M) | cd_seek_wedge f986 — benign |
shots/blue_lightening_bios_3000/frame_02700.png |
| Highlander | bios | PASS — intro FMV + title card ("HIGHLANDER (c)1994 Gaumont") | 25.5% | 98.2% | 1518 (1.85M/2.4M) | cd_seek_wedge f1755, f2716 — benign |
shots/highlander_bios_3000/frame_02700.png |
Screenshot paths are relative to the sweep scratchpad
(.../scratchpad/shots/); PPM originals sit beside the PNGs, logs in
.../scratchpad/logs/<tag>.log. Screenshots were visually inspected —
all show correct, recognizable FMV/title imagery, not garbage or black.
Why the watchdog lines are benign
Every cd_seek_wedge line above reports seek_starts == seek_dones
(no in-flight transfer; the drain counter is parked at the end of a
completed read) — the known benign "CD idle >5 s after a finished
transfer" class documented in CLAUDE.md (Myst intro case). Corroborated
against the motion timeline in each case:
- BrainDead 13: fires during its static title screen (67.6% nonblack,
screenshot
frame_01800.pngshows the castle title card). Motion resumes at win 036 (menu cursor blink) and win 044+ (attract). Thevideo_stallis the same static title screen — a real frontend shows a stable title card there, not a black screen. - Blue Lightning / Highlander: fire between FMV clips while motion continues in later windows (Blue Lightning ends at 94.9% nonblack with steady 15/60 motion; Highlander at 98.2%).
No gpu_pc_escape, dsp_pc_escape, gpu_wedge, or dsp_wedge lines in
any run.
Long-run probe — 15000 frames (~4 min 10 s emulated), both modes
Looking for late failures a 50-second run misses (wedges, black-outs, motion stopping). Screenshots every 1000 frames.
| Title | Mode | Verdict | Motion % | Watchdog lines | Notes |
|---|---|---|---|---|---|
| Dragon's Lair | hle | PASS | 61.5% | none | attract loops cleanly; recurring 2–3 s static card every ~3000 frames (same 30.9% nonblack each pass = attract-loop title card, periodic, not a stall) |
| Dragon's Lair | bios | PASS | 63.3% | none | frame 14000 screenshot = Dirk on the burning rope; FMV correct to the end |
| Space Ace | hle | PASS | 62.9% | none | clean to frame 15000 |
| Space Ace | bios | PASS | 63.0% | none | frame 14000 screenshot = Borf FMV; clean to frame 15000, RMS 1753 |
No late failure was found in any of the four long runs.
V2 CDI baseline — 600 frames, both modes ("before" evidence for the rip-repair track)
These 4 CDIs are known-damaged V2 rips (boot header region missing from
the rip itself — see project_cdi_v2_root_cause): the exact current
failure signatures, for comparison after any fix.
| Title | Mode | Verdict | Exact signature |
|---|---|---|---|
| Iron Soldier 2 | hle | FAIL — load rejected (retro_load_game returns false, exit 1) |
[CD-BOOTSTUB] Magic mismatch at +0x42 of session-2 track BIN (matched 1/32 bytes) → [CD-HLE] Boot stub extraction failed → [CD-HLE] Parked 68K on halt loop at $00000400 → [Virtual Jaguar] unsupported or invalid content format |
| Iron Soldier 2 | bios | FAIL — stuck in BIOS (loads; BIOS animation + jingle only) | [CD-BOOTSTUB] Magic mismatch at +0x42 ... (matched 1/32 bytes) → [CD-BOOTSTUB] CDIntfExtractBootStub failed at ~frame 420; run ends parked on the red JAGUAR logo (shots/ironsoldier2_bios_600/frame_00450.png), peak nonblack 21.1%, audio = BIOS jingle (RMS 1960, 240573 nonsilent — byte-identical across all four titles) |
| Myst demo | hle | FAIL — load rejected | same chain, matched 0/32 bytes |
| Myst demo | bios | FAIL — stuck in BIOS | CDIntfExtractBootStub failed; BIOS animation still cycling at frame 600 (motion 78.3%), RMS 1960 |
| Vid Grid | hle | FAIL — load rejected | distinct message: [CD-BOOTSTUB] Boot header region is zero-filled at +0x42 - this image is an incomplete / bad rip, not an unsupported format → same failure chain |
| Vid Grid | bios | FAIL — stuck in BIOS | zero-filled-header message → CDIntfExtractBootStub failed; parked on logo, RMS 1960 |
| World Tour Racing | hle | FAIL — load rejected | Magic mismatch at +0x42 ... (matched 0/32 bytes) → same chain |
| World Tour Racing | bios | FAIL — stuck in BIOS | CDIntfExtractBootStub failed; parked on logo, RMS 1960 |
The identical audio stats (240573/480000 nonsilent, RMS 1960) across all four bios runs are the CD BIOS boot jingle — direct evidence none of them ever leaves the BIOS.
Reproduction
ln -sfn "$JAGUAR_ROMS_PRIVATE" test/roms/private
DEVELOPER_DIR=/Library/Developer/CommandLineTools make -j8
cc -O2 -Wall -std=c99 -I. -I./libretro-common/include \
-o test/tools/cd_visual_verify test/tools/cd_visual_verify.c \
test/harness/harness.c -ldl -lm
VJ_HARNESS_LOG_INFO=1 ./test/tools/cd_visual_verify \
./virtualjaguar_libretro.dylib \
"test/roms/private/Jaguar CD/CDI/dragons_lair.cdi" \
--frames 3000 --outdir /tmp/shots --shot-every 300 \
--option virtualjaguar_cd_boot_mode=bios
# confirm in the log: [BOOT] CD game, mode=BIOS (BIOS image staged)
Conclusions
- Real-BIOS CD boot mode is not the cause of the user-reported FMV breakage — all 5 FMV titles verify clean under it, including 4-minute probes. The prior HLE-only sweep gap is closed.
- If users see FMV titles broken with good rips, the difference is
elsewhere (frontend, platform build, image variant) — the next
discriminator to collect from reports is the RetroArch log's
[CD-BOOTSTUB]lines and crash-watchdog signatures. - The 4 damaged V2 CDIs fail identically in both modes with loud, specific log signatures (above). Users with those rips get either an immediate load failure (HLE) or a BIOS logo that never advances (bios) — plausibly the actual source of "FMV titles broken, fine in BigPEmu" reports, since BigPEmu reads CDI headers differently.