Status

Headline

Current release is v0.9.3-ps1. Scenes validated under the project’s acceptance bar (pixel-perfect visuals plus synced SFX, signed off across every applicable variant – night, low-tide, holiday, raft-stage): 63 / 63. The current release line keeps every row in the live per-scene ledger signed off, all 126 high/low scene variants routed and timing-bearing, deterministic BOOTMODE scene selection logged and gated, the memory-region allocator promoted into mainline, and the extended-soak CD/VISITOR3-low stability fixes in place. BOOT allocations now seal after startup, CACHE allocations reuse long-lived resource storage through free-list/LRU paths, TRANSIENT scene allocations can be wiped between major scene loads, CD file searches are quiesced before directory walks, and VISITOR3-low no longer keeps the clean-relief stream window that caused a rare CACHE/libc fragmentation cliff. The chapter-select grid in the in-game Scene Explorer keeps the custom on-PS1-captured thumbnails for every one of the 63 scenes from v0.8.4-ps1 while streaming previews without a large paused-menu heap allocation. The public headless battle card is +0.1571% over target / 99.8440% target speed; the raw signed optimization matrix is -0.5598% / 100.5744%. The allocator validation branch records a R34 full matrix of 126/126 PASS with 0 BSODs, and the latest targeted W1/B2 plus VISITOR3 checkpoints keep the top allocator-era rows measured inside the allocator budget. The latest WALKSTUF1-high promotion presents and consumes frame0 before the measured loop, passing at 1462/1445, overrun 17, blocking/refill 35/5, and reads/due 36/6 while VISITOR3 high/low remain exact-flat under the same canary. The latest code-headroom pass makes JCPERF2 the default scene-end perf summary, compiles legacy JCPERF behind PS1_PERF_LEGACY_TRACE=1, and scopes ps1PerfEndScene() to -Os; the four-yellow canary stays exact-flat while that reporting symbol shrinks to 0xf4 bytes in the same PS-EXE bucket. The prior VISITOR3-low frame135 gap-D4 promotion keeps pack footprint, LBA, sectors, and the PS-EXE bucket fixed while improving VISITOR3-low 1071/1039 -> 1070/1039, overrun 32 -> 31, blocking 80 -> 78, and loop read time 106 -> 104. The prior W1-high cleanup-only fixed-layout screen-clip pass keeps file size, offsets, entry sizes, pack LBA, and the PS-EXE bucket fixed while clipping exact entries 14, 23, 43, 45, 47, 49, 51, 54, 56, 57, 67, 137, and 147; subset logical active payload drops 29339 -> 27643, 18160 cleanup pixels are removed with no draw-span removal, runtime restore bytes drop 525020 -> 510842, max restore bytes drop 12684 -> 7712, upload bytes drop 17185920 -> 17182720, and the five-yellow canary stays exact-flat. The prior VISITOR3-low fixed-layout previous-visible cleanup pass keeps file size, offsets, entry sizes, pack LBA, and the PS-EXE bucket fixed while changing 94 low-pack entries; logical active payload drops 425729 -> 405835, cleanup restore bytes drop 456786 -> 53456, runtime restore bytes drop 467962 -> 64632, max restore bytes drop 293544 -> 2962, upload bytes drop 18113920 -> 17494400, and the five-yellow canary stays exact-flat. The prior W1-high exact-entry preserve-entry-size screen-clip pass trims entries 58..61, drops subset logical active payload 17164 -> 9896, removes 4995 cleanup pixels plus 6052 draw pixels, and keeps the five-yellow canary exact-flat while preserving file size, entry sizes, pack LBA, and the PS-EXE bucket. The prior source-headroom pass caches the active foreground scene ID in hot scheduler paths, keeps the five-yellow canary exact-flat, and shrinks tracked foreground scheduler symbols while the PS-EXE bucket stays fixed. The prior BUILDING2-high no-decode trim-draw-tail headroom pass trims entries 74 and 78..82, drops active payload 548293 -> 539990, and keeps the five-yellow canary exact-flat while preserving file size and pack LBA. The prior BUILDING2-high no-decode trim-draw-tail speed promotion improves B2-high to 1341/1313. The prior VISITOR3-low preserve-entry-size screen-clip promotion keeps VISITOR3-low at 1071/1039 while preserving entry sizes and pack LBA. The prior W1-low preserve-entry-size screen-clip promotion keeps W1-low at 1470/1446 while improving hidden refill 4 -> 3 and preserving entry sizes and pack LBA. The prior BUILDING2 high preserve-entry-size screen-clip promotion improves B2-high to 1343/1312 while preserving entry sizes and pack LBA. The prior VISITOR3 high screen-clip headroom pass clips offscreen cleanup spans in entries 101 and 116, dropping active payload 437785 -> 436469 with exact-flat five-yellow timing. Release gates now require a nearby MEM_REGION_RATIONALE for every memAlloc call site. The prior accepted JOHNNY1 pack promotion compresses full-frame entries 1 and 50 behind a scene-local local-LZ sentinel, preserving pack footprint and the 217088 byte PS-EXE bucket while moving both tides to 1948/1945, overrun 3, blocking/refill 5, and target speed 99.85%. The accepted BUILDING2 low setup-segment promotions prime relative sectors 112..128 and 226..262 during setup without changing pack size/LBA, use a low-only 80 KiB clean-strip cap, raise the low window slack to 5, and add {141,153}, improving low to 1327/1318, overrun 9, blocking 47, reads 27, refill 0, and due 9. BUILDING2 high now layers guarded 271..287 plus 315..327 rows on top of the allocator-safe setup slices, then the previous-visible cleanup promotion moves it to 1343/1312, overrun 31, blocking/refill 50/17, and due 7; the entries 92/94/95 payload trim plus cleanup and no-decode trim-tail passes cut B2-high active payload 669408 -> 539990, and {185..197} plus {158..174} bank same-speed CD-pressure work under the new baseline. BUILDING4 low now carries the v971 local-LZ entry270 follow-up, the gap-8 dirty-upload band merge retune, and the 24 KiB stream-window promotion at 2847/2820, overrun 27, blocking/read time 32/252, and prefetch overrun 27, cutting active payload 807263 -> 799277 while preserving pack layout. The newest WALKSTUF1 baselines use allocator-safe targeted setup slices instead of full setup buffers: high now measures 1469/1440, blocking/refill/due 42/12/7, with loop reads/read time at 41/202 after keeping relative sectors 198..244, extending the second slice to 286..344, adding {149,165}, encoding frame 92 as previous-frame D4, adding {423,439}, {404,416}, {395,411}, and retargeted {411,423}, letting high tide use prepare-before-window scheduler ownership, and clipping exact entries 58..61 as the current tail/phase speed promotion; low replaces the old 197..243 plus 410..434 split with a single 238..344 retained CACHE setup segment after low-only 48 KiB clean-rect chunking, then adds the {91,107} first-boundary read group and a split TRANSIENT 344..350 setup edge, retargets setup to 244..350 plus split 179..185, adds {113,129}, and then adds {355,371} as same-speed read-work, then applies the preserve-entry-size screen-clip pass for hidden-refill headroom, and then layers the compact trim/retarget phase promotion. It now measures 1461/1447, blocking/refill/due 31/2/6, with loop reads/read time at 22/117. Both paths keep pack LBA/sectors and the PS-EXE bucket fixed while reducing CD pressure. VISITOR3 now keeps only its tiny stage1 prefetch frame buffer plus bounded clean-relief stream windows: high uses an 80 KiB window with terminal overread trimming, merged setup segment 203..262, frame 139 raw-gap relocation, frames 56 and 57 raw in that paid gap with a 56 KiB tight-refill cap, high clean strips capped at 64 KiB, and the early retained setup edge 40..47. High now measures 1065/1046, overrun 19, blocking/refill 34/1, reads/due 8/2; low uses a 16 KiB slack-5 window plus a third retained setup segment extended to 206..232, with frame 138 raw relocated into that paid gap, then adds the 16..32, 72..88, and 88..104 retained read-group rows plus the four-VBlank slack-knee guard, then adds a one-VBlank low-tide phase offset and widened setup segment 38..79. It now measures 1062/1045, overrun 17, blocking/refill 38/0, reads/due 12/6. Both rows moved out of red and the orange band is now empty without reintroducing the clean-rect allocation failure. BUILDING4 high now primes relative sectors 264..288 during setup, improving 2847/2816 -> 2843/2816, overrun 31 -> 27, blocking/refill 36/32 -> 34/30, and moving that row into green. VISITOR5 high/low both remain green on the matching 30..46 retained-read shape. WALKSTUF1 high/low now measure 1469/1440 and 1461/1447; the latest high frame92 D4 stack cuts high overrun 34 -> 31 without moving layout, the latest high scheduler/phase/tail rows cut blocking/due to 42/7 and overrun to 29, while low keeps the retained setup stack and moves into green through compact trim/retarget phase work. BUILDING2 low now keeps the accepted 218..229 slack-8 retained-read row plus v739 draw-tail trimming, then primes relative sectors 112..128 and 226..262 during setup, with the allocator-era matrix at 1327/1318, overrun 9, blocking/refill 47/0, reads 27, and due 9. BUILDING2 high currently measures 1341/1313, overrun 28, blocking 47, refill overrun 14, and due misses 7; the allocator-safe setup slices plus the 83..95, {158..174}, guarded 271..287, 315..327, {185..197}, and previous-visible cleanup promotion leave reads/read time at 44/194, and same-speed offscreen clips now reduce its runtime rows/spans/pixels to 18030/105645/446246, and the v877/v879/v880 preserve-offset frame172/frame171/frame96 trims plus the entries 92/94/95 follow-up reduce active payload 674798 -> 663590; the cleanup and no-decode trim-tail passes drop it further to 539990 without moving pack layout. The live ledger is at /scenes/; the per-scene workflow that drives the bar is in docs/ps1/scene-status.md.

The narrative below is per-component, not per-scene. The regtest harness (config/ps1/regtest-scenes.txt + scripts/run-regtest.sh) is the source of truth for “does this still build and boot.” Human visual + audible signoff is the source of truth for “is this scene done.”

Per-component status

Component	Status	Notes
Build system (Docker + CMake + mkpsxiso)	Complete	config/ps1/Dockerfile.ps1 on linux/amd64; PSn00bSDK 0.24; scripts/rebuild-and-let-run.sh is the one-command entry.
CD-ROM I/O (cdrom_ps1.c, ~2,570 lines)	Complete	Reimplements fopen/fread against CD sectors. CdSearchFile / CdRead / CdSync integration. The rest of resource.c is unchanged from upstream.
Renderer (graphics_ps1.c, ~4,900+ lines)	Complete	4-bit indexed sprite format (indexedPixels), palette LUT compositing, 4-pixel unrolled inner loop, opaque sprite fast-path, dirty-rect background restore (~80-95% reduction in per-frame data movement), black-backdrop temporal cleanup for full-screen scenes like JOHNNY 1, and grForceFullRedrawNextFrame for pause-menu resume. FntFlush is empirically broken in the scene-runtime context – do not regress on-screen text to it.
Audio (sound_ps1.c, ~445 lines) + SPU	Working	All 23 VAG SFX preloaded into SPU RAM at boot. Round-robin over 8 voices. Captured 0xC051 PLAY_SAMPLE events ship in the FG2 pack and fire from foreground_pilot.c with a 3-frame key-on delay. Mute writes the SPU master-volume registers directly because SpuSetCommonMasterVolume is not honored by DuckStation HLE. The VAG encoder (scripts/wav2vag.py) was extensively debugged during the v0.3.6-ps1 milestone (commit 355227fa); see that commit for the full bug list.
Input (events_ps1.c + src/platform/ps1/spi.c)	Complete	Direct SPI driver, timer-2 + SIO0 IRQ at 250 Hz, lifted from spicyjpeg’s pads example. The BIOS pad path (InitPAD/StartPAD) is unusable in PSn00bSDK 0.24 + DuckStation. Poll TX is tx_len=5, not 4 – DuckStation only delivers button bytes when the full 5-byte sequence comes from the TX buffer.
Closed captions (src/ps1_captions.{c,h})	Working	On/off via Pause -> Accessibility -> Captions. Dark band at the bottom of the frame for ~5 seconds at scene start with descriptive subtitle text. Glyph atlas shared with the pause menu. Caption corpus from the upstream closed_captions branch of jc_reborn; the original sequential ADS-tag map had ~20 mismatches and was re-audited (docs/ps1/caption-audit-2026-04-26.yaml). HIGH-confidence matches dominate; LOW-confidence slots remain on STAND idles and a few VISITOR / WALKSTUF edges.
Holidays (36 of them, code-generated)	Working	Holiday emblem sprite sheet packed into the PS1 holiday overlay. Selectable via Pause -> World Options -> Holidays and BOOTMODE.TXT. Generation is offline; design notes in docs/ps1/holidays-expansion-design.md.
Story-loop walks (walk_pilot.c, walk_render.c)	Working	Johnny walks between scene endpoints using Sierra’s original walk_data.h routes instead of teleporting. The runtime restores from a persistent tight clean buffer, keeps waves moving, re-stamps holiday overlays behind Johnny during walk frames, and covers Johnny behind the palm trunk/leaves. v0.7.2 adds a backdrop-key guard so walks only run when the next scene matches the previous rendered tide, raft, night, holiday, and island X/Y state; v0.8.0 adds a clean-rect retry path; v0.8.1 counts wave-band/split-rect clean pressure before allocation.
Freeplay/debug mode (scene_freeplay.c)	Working	v0.5.0-ps1 promotes direct-control Johnny: D-pad/analog walking, L2/R2 speed modifiers, fishing, immediate R1+D-pad world toggles, gag/visitor debug catalogs, sound test, Select clear-screen rebuild, frog-clock loading transitions, and a no-allocation steady-state frame loop.
Pause menu (pause_menu.c)	Complete	Start opens the overlay mid-scene; custom embedded 8x8 ASCII font (because FntFlush is broken in scene context); POLY_F4 dim quad + panel quads. Twelve sub-screens: Scene Set, Scene Explorer, Freeplay Options, Controls, World Options, Holidays, Set Island Position, Accessibility, Sound Test, System, Set Time/Date, Set RNG Seed.
Memcard persistence (memcard.c)	Working / expanding	Pause-menu choices persist to bu00: block 0. Save/load wired; restore-on-boot wired. v6 saves persist holiday mode separately from manual holiday id. Broader menu-option persistence remains future work.
Telemetry / debug overlay	Complete	Five-panel left-edge overlay; gated by grPs1TelemetryEnabled in graphics_ps1.c with row writes scattered through src/ads/ads.c and helpers like grPs1SetLastBmpTelemetry. Off by default in release builds. Decoded out of captured frames by scripts/decode-ps1-bars.py into telemetry.json during regtest. The historical visual-debug substrate when TTY printf was unsafe; still the right tool for per-frame state since text I/O can perturb timing.
Perf instrumentation (ps1_perf.c)	Complete	Level-gated JCPERF2 TTY summary by default; legacy JCPERF can be re-enabled with PS1_PERF_LEGACY_TRACE=1. Levels: OFF, SUMMARY, DETAIL, DEBUG via ps1PerfSetLevel. Off in normal operation.
TTY printf	Reliable	Restored 2026-04-25 on PSn00bSDK 0.24 + DuckStation through bounded vprintf plus DuckStation TTY/file logging. Gated BOOTMODE probes (printf-test, logtest). Must not be called per-frame – text I/O alters timing.
Regtest harness	Working	config/ps1/regtest-scenes.txt + scripts/run-regtest.sh drive a headless DuckStation pass. Source of truth for “boots and renders something.” Not the source of truth for “looks right” – that’s still human signoff.
Host capture pipeline	Working	scripts/capture-host-scene.sh runs the desktop build under controlled boot state; emits high/low PNG frames, frame-meta.json, sound-events.jsonl. scripts/export-scene-foreground-pilot.sh now defaults new scene bring-up to normal/far-left/far-right foreground-only multi-view stitching before compiling FG2.
CD packaging (mkpsxiso)	Complete	config/ps1/cd_layout.xml. Routed scenes contribute high-tide + low-tide pack entries; the full FG2 corpus is ~343 MB, of which a selected subset rides on the disc. Build outputs jcreborn.bin + jcreborn.cue. CD image at v0.9.3-ps1 is ~79 MB; PS-EXE is ~208 KiB after legacy ADS/TTM/FG1 paths were stripped from the linker pass.

What’s currently broken

Honest list, narrowed to specifics:

• Post-validation regressions. Scene coverage is complete at 63/63, but future bugfixes can still be needed when broader play, new features, or performance changes disturb a signed-off path. v0.8.1 fixed the latest random-run clean-rect pressure freeze; the scene ledger remains the acceptance source after each fix.
• Memcard fresh-card edge cases. Save / load works on a card that already holds the project’s block. Behavior on a freshly formatted card or a card with a corrupted block is still being shaken out.
• SpuSetCommonMasterVolume HLE divergence. Worked around with direct register writes; the divergence itself is a DuckStation HLE quirk, not a project bug, but it does mean audio-on-real- hardware needs verification before any release calls itself hardware-confirmed.
• Real hardware soak. Smoke-tested on an SCPH-7501 via the TonyHax softmod path — the disc boots and runs. Long-term observations on real hardware (extended idle, real SPU register behavior, real CD seek timing under random-position scene picks) are still on the wishlist. The SpuSetCommonMasterVolume HLE divergence above is a concrete reason to keep treating hardware audio as unverified until the soak loop covers it.
• Older count-based status models (25/63, 27/63, 60/63, 63/63) from prior validation eras are retired, not current. They are preserved in current-status.md for searchability and demoted explicitly. Do not cite them as current progress.

If a specific bug isn’t called out here, it lives either in the GitHub issue tracker on huntergdavis/johnny-castaway-ps1 or in the dated worklogs under /devlog/. The regtest harness output is the SoT for build-and-boot regressions.

What’s currently in motion

Pulled from the live narrative in docs/ps1/current-status.md:

• Optimization after full validation. With all 63 scenes signed off, the next focus is preserving pixel-perfect playback while improving speed, loading, memory pressure, and release polish. v0.9.3-ps1 has the public-capped headless baseline at 99.8% target speed across all 126 timing-bearing rows after the VISITOR3, WALKSTUF1, MARY3, BUILDING, ACTIVITY9, JOHNNY1, and WALKSTUF3 promotion arc. The bring-up loop remains in docs/ps1/development-workflow.md.
• Scene-by-scene FG2 routing. All 63 scenes have generated high-tide and low-tide FG2 packs sitting in the corpus; routing them onto the CD image is selective during bring-up so the image stays manageable.
• Memcard hardening. Closing the fresh-card and corrupted- block edge cases.
• Persisted-options continuation. v0.7.1 shipped the holiday-mode slice: AUTO DATE:ORIG4 is now the fresh/no-card default and memory-card schema v6 stores automatic policy separately from manual holiday id. Captions, Scene Set, tide, raft, island-position, freeplay/perf preferences, and menu cursor defaults remain future work.
• Caption audit follow-through. Two layers, with one done in v0.8.4-ps1 and one still open. Done: the website’s description of which caption belongs to which scene was reconciled against the on-PS1 packs during the chapter-select grind — per-scene index.md titles + bodies, scenes.yml notes, scene-status.md Notes column, and the in-game Scene Explorer display strings. The runtime evidence overturned a substantial fraction of the audit’s confidence ratings, including two HIGH-rated FISHING entries (FISHING 2 boot, FISHING 3 crab) that were exact text-pack mismatches and the VISITOR 5 NO_MATCH candidate that turned out to be a clear unambiguous gag. Open: the runtime captionSceneMap[] array in src/platform/ps1/ps1_captions.c was not changed in v0.8.4-ps1; the on-screen caption band continues to play whatever caption the original audit picked, even on rows where the website now disagrees with that pick. Repointing the runtime mapping (or blanking the rows where text and pack can’t be reconciled) is a future-release item. The post-validation runtime corrections section lists the named mismaps.
• fgpilot -> “PS1 scene playback” naming migration. The internal code-name in foreground_pilot.c and friends remains fgpilot; the public-facing name is moving to “PS1 scene playback.” The migration plan lives in docs/ps1/ps1-branch-cleanup-plan.yaml under fgpilot_naming_migration_plan.
• Milestone release cadence. v0.9.2-ps1 hardens the long-run CD and VISITOR3-low paths after multiday soak testing. v0.9.1-ps1 raises CACHE headroom and improves BSOD display for longevity testing. v0.9.0-ps1 marks all 126 routed high/low scene variants green at 99%+ target speed. v0.8.16-ps1 promotes the memory-region allocator: BOOT/CACHE/TRANSIENT lifetimes, scene-boundary transient wipes, CACHE free-list/LRU reuse, generated pack-header metrics, and allocator rationale gates. v0.8.15-ps1 promotes the WALKSTUF1 high setup-resident CD row. v0.8.14-ps1 promotes JOHNNY1 high/low local-LZ full-frame compression. v0.8.9-ps1 promotes the current mainline optimization set: VISITOR5 low reuses the green 30..46 retained-read shape, BUILDING2 low carries the 218..229 slack8 row plus v739 draw-tail trim, and WALKSTUF1 low carries no-shift frame 51/49/47/61/62/58/45/37/35 in-place payload reductions. v0.8.8-ps1 promotes the VISITOR5 high 30..46 retained-read group on top of the v0.8.7 stability hardening. v0.8.7-ps1 is the deterministic BOOTMODE scene selection and Scene Explorer preview stability point release on top of the v0.8.6 performance follow-through: direct-scene boots are logged and gated, Suzy backdrop cleanup restores scene state, and Scene Explorer thumbnails stream without a large heap allocation. v0.8.6-ps1 is the WALKSTUF1 / VISITOR3 setup-segment compaction follow-through point release on top of v0.8.5’s full 126-row baseline (WALKSTUF1 low gap6-prefix + slack-guard, WALKSTUF1 high window-prefetch / slack4 guard, VISITOR3 high/low setup-segment resident copies for frames 131 / 128). v0.8.5-ps1 is the full 126-row headless performance matrix release. v0.8.4-ps1 is the chapter-select thumbnails content release — 63 custom on-PS1-captured grid slots and a scene-page reconciliation pass against the actual on-PS1 packs. v0.8.3-ps1 is the WALKSTUF1 compact foreground performance pass. v0.8.2-ps1 is the VISITOR3 guarded-read performance pass. v0.8.1-ps1 is the clean-rect pressure stability point release. v0.8.0-ps1 is the complete-scene performance baseline. v0.7.2-ps1 fixes stale-backdrop story-loop walks. v0.7.1-ps1 adds the original-four auto holiday default and persisted holiday mode. v0.7.0-ps1 validates all 63 scenes and moves the project into bugfix/optimization mode. v0.6.10-ps1 validates MARY 5 after the NORAFT raft clamp and FIRST full-wipe transition fix. v0.6.8-ps1 validates MARY 2 after the wide multi-view stitch and fish thought-bubble repair. v0.6.6-ps1 removes the old FISHING7/FISHING8 runtime pins after far-left recapture proved both packs are random-position safe. v0.5.0-ps1 is the freeplay/debug release. v0.4.20-ps1 was the walking-loop stability release. v0.3.9-ps1 was the fishing3-loop-stability release; v0.3.6-ps1 was fishing1 with full SFX across all variants.

For dated context on any of the above, see /devlog/.

Related pages

• About: Method — the canonical methodology essay this status page tracks the implementation of.
• About: History — how the components above got to “complete” or “in motion.”
• About: dev environment — the workflow photograph behind every row in the component-level table.
• Scene ledger — the visual signoff bar; the per-scene status the table summarizes.
• Performance battle card — the second ledger; per-scene/tide timing for every component listed under “performance” above.
• Releases — short notes on every tagged version named in the milestone-cadence list above.
• Glossary — vocabulary anchor for FntFlush, clean-rect, fgpilot, BOOTMODE.TXT, and the rest of the terms this page uses without scaffolding.