Research and Retired Work

Research and Retired Work

Current feature register · Actionable research · Upstream work · Performance evidence

Reviewed: November 1, 2026. Fork source: b7140848 (full SHA b71408480a3a979a9c584663e1d1f400904c8bb7). Latest stable checked: v0.8.136-0-g458730bf5373, September 30, source 458730bf5373f1f0d027450fef7443e3af7a3ae1. GitHub state, branch recommendation, source presence and release availability are separate facts. See current status and release versus main.

This register explains work that is removed, preserved on an unmerged branch, or still held for qualification. It prevents a merged predecessor, an attractive PR title, a saved prerelease or a component result from becoming a current product promise. The linked current PR bodies contain detailed branch code, failures, identities and bounded remaining work.

Removed and held product directions

Work	GitHub state / disposition at review	What exists and what remains	Pinned main / checked stable
Shader Browser #237 / #238 / #239	Merged, then removed by #268	Historical experiment; UI, product files and SQLite/override integration removed. Old user databases may remain unused on disk.	Absent / absent; removal predates September 30 stable.
Workbench / Asset Browser #241 / #259 / #260 / #242	#241/#259/#260 closed unmerged; #242 closed, superseded by #259	Preserved branch research. Synthetic/captured previews do not promise faithful complete-frame replay or a general asset importer.	Absent / absent; no current UI promise.
Neural / DLSS scaffold #262	Open / HOLD	SDK-neutral presentation adapter and restart-latched settings scaffold only. Genuine neural evaluation, motion/depth generation and validated rendering integration remain unimplemented. Windows SDK/proxy and native GPU qualification missing.	Absent / absent; no delivered DLSS pipeline.
Depth alias runtime consumer #290	Open / HOLD	Merged foundations are separate. Guarded alias consumer retains historical Conker benefit, but restored scene/progress, reset/load and surface-oracle gaps remain; original six-arm campaign stopped.	Consumer absent / absent; later foundations main only.
Small color upload consumer #291	Open / HOLD; child of #290	Linux 1x 32×32 swizzled color branch; display upload-before-finish repair present. Parent lifecycle/oracle and current qualification gates remain.	Consumer absent / absent; not a general upload path.
Linear voice resampler #189	Open / HOLD	Optional experimental selector remains on branch. Current main still uses SRC_SINC_FASTEST; mono sinc and ADPCM mapping improvements are separate delivered work.	Absent / absent; ordinary path remains sinc.
Attenuation lookup #193	Closed unmerged	Measured mixed research: eight-run Deck report gives +0.70% display-write cadence and 0.62% lower process CPU, but p99 6.32% worse in both orders. Preserve the tail finding; no accepted landing.	Absent / absent; current attenuation still uses powf().
Exact stereo ADPCM transitions #195	Closed unmerged / SKIPPED; HOLD evidence	Varied-input component gains did not supply an integrated win: Deck Morrowind used 2.21% more CPU. Tested layout probe did not recover cost; separate row8/SSE2 redesign remains unimplemented research.	Absent / absent; decoded-block reuse in main is distinct.

The Browser removal is independently consistent with product-path absence in both pinned trees. Shader Workbench branches remain useful engineering records, but no browser, database integration, override runtime or importer should appear in a current menu guide. Retained user files are not an active product database.

Audio source checks: main voice processing still constructs SRC_SINC_FASTEST resamplers and computes attenuation using powf(). Existing mono sinc, decoded-block reuse, callback-local voice format and ADPCM mapping work must keep their separate delivered status; closing #193/#195 does not undo those independent landings. Audio architecture and the feature register describe current behavior.

GP trace, replay and native-filter stack

All five PRs are open, unmerged and HOLD, absent from pinned main and the checked stable release. #376 retains the archived pre-isolation #375 lineage (archive/pr375-pre-isolation-20261101), not the current merged #375 boundary. Integrate/review the required stack and qualify the combined result; a useful child does not silently qualify its parents.

PR	GitHub state / recommendation	Current head	Bounded implementation and evidence
#376	Open / HOLD	bf9b88882641410c9b55aac2ca60ed15aaa2ec5c	Optional GP/EP host-wall trace; one PGR2 diagnostic assigns 94.67% of summed DSP elapsed intervals to GP. This is not CPU attribution or speedup. Observer cost/parity and stack integration remain unqualified.
#377	Open / HOLD	f8a029810a03c43e35768f316362d29e30c90c44	Non-destructive JIT peripheral export; real-backend checks repair destructive DMA-poll consumption and retain the 24-bit mask. Complete-VM snapshot/engine/PCM and timing qualification remain.
#378	Open / HOLD	f47da478581bf274b6da01351d36939faa96e11b	Opt-in complete GP capture and real-JIT replay. One captured frame matches 44 calls, 57 transactions and complete poststate on Deck/build host. Two failed diagnostic setups remain; no hardware/audio or performance proof.
#379	Open / HOLD	3a37ba573daf8913f86526eb6bb20c6856197255	Opt-in recognized-program native GP filters, default off with conservative JIT fallback. Supported frozen frame reduces FFI calls 44→25 with complete-state/I/O agreement; fresh N3/N4=1 frame falls back. CPU-clock campaign failed validation (nine zero thread-CPU samples); no accepted speedup.
#380	Open / HOLD	cd61010a222407b7047908084b9ed70f9298c059	Narrow paired N3=N4=1 admission above #379, retaining pair 2 and fallback. Frozen pair-1 frame reduces FFI calls 44→27; scene-local native coverage observed. Fresh pair 9 remains fallback. Combined-stack cost, PCM/game, lifecycle and XISO timings missing.

Full-state replay and fewer FFI calls are correctness/coverage results, not CPU savings, audible parity, guest cadence or host-presented FPS. Current GP timing failures remain failed/ineligible. Physical Xbox parity and complete-VM PCM/lifecycle checks remain required; diagnostic captures/manifests are private operator artifacts linked from the owning emulator PRs, not publicly reproducible evidence or files merged into main.

CPU, graphics and identification research queue

See the current fork action table for exact heads and separate GitHub/recommendation fields:

• #317: Open, SKIPPED/HOLD. Recover original self-page/SmartXX assets; do not rerun unchanged zero-exposure PGR2/Conker scenes.
• #318: Closed unmerged, SKIPPED/HOLD. Retain the single Deck pair’s +2.305% guest cadence and mixed tails, including the worse isolated maximum; suitable hardware/workload may justify reopening.
• #319: Open/HOLD. All original game attempts collected; diagnose stale B3/B4 frame traces and physically qualify the new border-color repair. Completed component/XISO work does not qualify those game attempts.
• #320: Open/HOLD. Branch admission defects repaired; native ARM ordering and generated-code qualification missing.
• #321: Open/HOLD. Actual conversion, comparison, rounding/exception and default-on-without-rollback defects remain; softfloat/admission correctness precedes activation.
• #351: Open with QUALIFY title and HOLD recommendation. Current qtest/build/integration checks pass; recover original launch/probe identities and finish native/resource and XISO qualification. Optional identification device is absent by default; flash, LCD, switches and complete physical settings behavior remain unimplemented.

Upstream history and maintenance

The upstream register retains all 26 September pages under historical headings and separately records current state/head identity. #2956 closed for still-open #2988; #3000 merged and its equivalent fork target-selection mechanism is independently present in stable/main; #3014 closed with no verified closing reason. New #2965 and #3058 heads require refreshed source review.

This documentation refresh ran no new emulator benchmark, XISO or hardware campaign. Preserve historical recommendations and failures with their tested identities. Change this page when a branch actually merges, closes, changes its recommendation, or loses source presence; update release availability independently.