psxrecomp/docs/PARITY_CAMPAIGN.md

PlayStation enhancement parity

Owner-approved scope, 2026-10-06. Tracking: beads-eio.3.277; shared execution work: beads-wuhs. Implementation is in progress. Source presence, a successful build, a short gameplay check, and owner acceptance are distinct milestones.

Acceptance target

• Enhanced OpenGL is the normal renderer, with 1080p internal resolution.
• Native scene interpolation targets Display refresh. Temporal image blending alone does not meet this requirement. HUD, FMV, menus, pauses and transitions need appropriate presentation behavior without advancing gameplay twice.
• PGXP, stable world texture filtering and perspective correction apply where useful. Preserve pixel-art/UI treatment and fix seams, jitter and texel bleed.
• Adaptive view covers 4:3 through at least 32:9: world geometry, culling and activation, backgrounds, textures and HUD layout. Wider displays may stretch.
• Extend distance/fog and replace unnecessary subdivision when applicable; preserve visible content and required game behavior.
• Use resident assets, native decompression and bulk transfers where measured loads justify them. Initial startup, transitions and repeat visits all count.
• Qualified HLE is selected in normal builds; a functioning LLE reference is separately buildable. Compare required ABI/results/effects, and measure useful speed. See HLE_EXECUTION.md.
• Preserve and validate existing multiplayer enhancements. Do not add multiplayer to single-player titles as part of this campaign.
• Deliver reproducible Windows ZIP and Linux AppImage candidates for the primary disc variant. Other variants are explicit follow-up work.
• Owner playtesting supplies final acceptance. No title is marked complete merely because its settings expose these options.

Starting inventory and remaining work

This records implementation evidence found during the initial repository audit, not full-game qualification. Framework baseline is 7b253941 (CODEGEN 18).

Title	Existing evidence / starting point	Work still to qualify
Tomba!	3c8d36e; custom native renderer, shared resident kit	Native interpolation replacing blending, broad resident coverage, full adaptive view and packages
Tomba! 2	5fc54c8; shared seamless USA loading	Native interpolation replacing blending, view/residency coverage and packages; Italian variant later
Mega Man X4	51f5f44; earlier widescreen/blending	Native interpolation, current renderer/filtering, load path and view coverage
Mega Man X5	c8fde08; earlier widescreen/blending	Native interpolation, current renderer/filtering, load path and view coverage
Mega Man X6	master 3b4b5dd; native interpolation; resident/mod branch 3bf0205	Integrate resident work with current framework; all applicable defaults, views and packages
Ape Escape	release a8e219a; native interpolation, private native packet producer	Shared build-selected HLE, gameplay interpolation qualification, startup/load service and packages
Vigilante 8	fidelity/packaging checkpoint 3591161; OpenBIOS Ski Resort gameplay, adaptive 4:3–32:9, PGXP, 1080p and stable filtering checked	Native interpolation, shader batching performance, load throughput, Linux package and broader gameplay; donor content excluded
Vigilante 8: 2nd Offense	e1133d10; native replay, PGXP, filtering, resident load-floor improvement (~307 to 45 frames in one case)	Backgrounds, subdivision/distance and remaining performance cases; full validation
Tsumu	9fa07da; English localization, older software/4:3 baseline	Applicable enhanced rendering, presentation, adaptive view and load audit
Tony Hawk's Pro Skater 2	6890a0da; native replay, extended distance, existing multiplayer	Wider visibility and distance limits, single/split/network play checks, load audit
Jersey Devil	a3958a3; native replay across banks	Full visual/view/residency matrix and current candidate packages
Spider-Man	66412b4; native replay and PGXP	Full view/texture/load coverage and current candidate packages
Spider-Man 2	60c8e80; PGXP, motion-adaptive image blending	Native interpolation under implementation, then view/load and package qualification
MediEvil	enhancement branch 84202de; native replay, PGXP, filtering, 3x distance, subdivision bypass	Resident loading adapter, full view and gameplay coverage, packages
MediEvil II	master 6816fc0; quality reference 32ea083; 5x/1080p, resident PP20/levels	Preserve reference features and validate full game; one measured load improved 8.30 to 1.38 seconds

Vigilante 8 content parking

The owner abandoned cross-game assets. These local branches retain that work:

Branch	Preserved checkpoint
archive/v8-v10-content-20261006	8a68555
archive/v8-n64-arena-20261006	ea3a006
archive/v8-renderer-and-media-20261006	8169c75

The earlier sequel archive at f3e4ab1 is also retained. The active feat/v8-fidelity-parity-20261006 branch removes donor importers and converter tests. Fidelity, rendering performance and loading speed are its only focus.

Shared work and validation receipts

The framework now has build-selected implementation families, implementation manifests, snapshot/network execution compatibility, and a stateless native GTE entry for bulk services. It is not a measured speedup by itself. Ape's packet producer is the first integration consumer.

The initial six targeted CTests pass: execution profile linking and source identity, reference/enhanced GTE contracts, reference/enhanced snapshot handoff and atomic rejection, and network content gating. Production main/GTE/snapshot translation units compile, including production netplay with networking enabled. Ape, Spider-Man 2 and Vigilante 8 Windows runtime candidates have linked; useful title HLE performance comparisons are still required.

Baseline framework tests also exposed three existing stale fixtures: gl_texture_filter_test, gpu_gp0_history_guard_test, and savestate_status_protocol_test. Their wiring is repaired and all three now pass, including the hidden real OpenGL filtering probe. A complete suite run against the new worktree remains outstanding.

Execution and promotion

The owner authorized one Sol worker per title in bounded waves. First wave: Ape Escape, Spider-Man 2, and Vigilante 8. Shared changes are owned by the coordinator. At most two compile jobs and one game process run at a time; validation uses isolated saves and ports. Preserve personal sessions and saves.

For each title record disc identity, framework/game revisions, actual applied features, rejected cases, screenshots and gameplay paths, load/frame metrics, build profile and manifest, package hashes, and owner feedback. Shared mechanisms belong in psxrecomp; disc addresses and asset contracts belong in the game. A missing or inapplicable feature needs an explicit reason.

Candidate branches and local commits are authorized. Public source pushes, merges and releases are not part of the current authorization. Beads stays open through owner acceptance. The central Dolt remote currently fails to push due to a missing remote data ref; local issue updates remain durable.

Use tools/qualify_render_passes.py during a known playable scene, with replay verification enabled. It records before/after counters, the supplied executable hash and its execution manifest. Zero replay, image blending, disabled replay, counter resets or verification failures cannot produce a passing sample. This is one evidence window; visual coverage, load speed and owner acceptance remain separate checks. Use tools/load_probe.py for passive load-window measurements.

Run correctness and performance windows separately. PSX_RENDER_PASS_VERIFY=1 adds synchronous GPU readbacks plus state hashing. After recording correctness, relaunch without it and use --purpose performance. The tool rejects a performance sample with active verification. Spider-Man 2's initial verified replay took 28-35 ms; verification off measured 19.531 ms, of which 19.176 ms was guest draw. That evidence points to title geometry processing rather than shared GPU checkpoint/capture as the remaining throughput bottleneck.

The shared presenter now honors HOLD for native phase generations instead of crossfading neighboring phase images. Phase selection tests and the real OpenGL probe pass, including red/blue native images that must stay distinct under HOLD while explicit blend mode still mixes them. Title requalification on this correction remains required.

The expanded GL probe now passes 509 checks. Shared commit 7c9e13ec carries nearest/bilinear/stable selection per primitive, preserving sharp untracked UI while batching filtered world geometry. Across all three texture depths, native and 5x rendering and native-wide margins match the original filter-boundary submission exactly. Its opaque fixture goes from eight batches to one; mask checking stays at eight and semitransparency at nine. Vigilante 8's measured filter-transition pressure motivated this work; game speed is still being compared on an isolated shader-only framework checkout.

The campaign framework checkpoint is 04f871f0. It also includes:

• 8fe08921: pre-materialized generated reference bodies build through include wrappers, avoiding duplicate Ninja rules without maintaining copied LLE code. The profile fixture checks both profiles, source-specific flags, local includes, automatic changed-body identity and relocation.
• 88730937: render replay refuses SIO/CD/MDEC/SPU MMIO reads before their uncheckpointed side effects and suppresses I_STAT's SIO tick. The real transaction fixture verifies rollback/recovery; production memory, replay and debug translation units compile. This is separate from correcting Ape's unsuitable broad draw boundary, which reached a real CD command/poll loop.
• 04f871f0: generation- and address-guarded packet precision capture/restore, including derived coordinate flags, projection and absent shadows. The new fixture and existing PGXP suite pass. Ape's smaller draw span uses this to preserve recorded prefix/tail packets only when packet ownership and link layout remain compatible; that title candidate still needs gameplay checks.

Vigilante 8's earlier Windows candidate was staged twice byte-identically: Vigilante8PSXRecomp-qualified-candidate-windows-x64.zip, SHA-256 060d8ef2195ca43162564870675b6b15ef2e49147d758e7cebf228cb62b08558. It excludes retail disc/BIOS and player saves, retains audited native AOT pairs, and passes execution identity/catalog/ZIP checks. It predates the shader batching comparison and is not final parity acceptance. No Linux candidate or measured V8 load-speed gain is claimed.

The shader-only V8 comparison did eliminate filter-change splits and reduced CPU flush time, but it did not establish a GPU/FPS improvement. Even the first window's similar primitive counts hid different textured fractions; later windows and camera images diverged. Host precision/mod-renderer state around the saved checkpoint needs investigation. The earlier b0/d8 title candidate remains qualified, with the 7c executable and observations retained separately. Use a fixed-work renderer probe before drawing a performance conclusion.

Shared regular-file identity hashing now uses Windows BCrypt or available Linux OpenSSL, with the portable SHA-256 implementation as fallback. Failed native initialization, update or finalization restarts the entire file; I/O failure leaves the caller's digest untouched. The existing C SHA context and CHD decoded-sector fingerprint path are unchanged. This host utility preserves exact identity in both execution profiles; it is not a guest HLE selection or a shortcut that skips content verification.

The Windows native/portable suites, mod runtime, resident-loading and PGXP session CTests pass (5/5). Linux OpenSSL and portable builds pass 68 and 67 checks respectively, including Unicode paths, SHA block/read-chunk boundaries, known vectors, partially consumed native failures and unreadable inputs. An isolated interleaved six-sample Windows measurement on V8's 257,200,608-byte data track matches Python's SHA-256 in every sample: portable median 0.656533 s, native median 0.130985 s (5.012x for hashing). Receipts are private under _build-parity-render-perf-20261006/{validate_hash,linux-hash-validation,benchmark_hash}.json. This saves about half a second for that operation, not the entire multi-second launch delay observed on the busy host. Overall startup still needs attribution.