[Felipebmaitan]

What

Custom submixers currently use one sample buffer for inputs, output ports and Gain. Add 1-4 per-voice buffers for custom submixers with zero modules or one Gain stage: inputs feed matching buffers, each grain clears unwritten samples, ports select their configured source, and Gain overwrites its selected destination. Broader chains and module routing stay unsupported until their processing-order contract is established.

Correct Gain's 37 unclamped levels, previous-level i/N ramps, equal-gain fast path, steady zero clearing and steady unity copying. Invalid Gain positions return 0x804A835A and stop linked parameter processing. Contracts follow the PS5 12.70 module report and later findings; KytyPS5 corroborates port selection. No new console measurement is claimed.

Unverified Gain initialization/lifecycle, callback setup/route changes and nonfinite integer rendering retain their runtime refusals and debt entries. Sampler Gain, broader custom chains and channel conversion remain outside this milestone. Tests exercise supported operations rather than requiring unimplemented cases to keep throwing.

Dean's requested follow-up is implemented at c40ba98945e845b950ecca7819c472942757ad30. Merged main 9b520147fbeddff4357d6e8f92e5235a16342171 in 0c3f104067239abd5358b8d1992295f9da7b27ee; then removed the unsupported exception assertions. CustomSafety keeps ordinary custom/upstream callbacks, explicitly configured Gain, route changes before rendering, callback counts, port metadata and literal output checks in both libraries. Gain keeps finite integer controls and its float/numeric/ramp coverage. Ngs2Custom's supported coverage is retained.

The earlier proposed queue's #9926, #10127 and #10410 are now in main, including #10410's PCM-fixture lifetime repair. The merge preserves their PLAY/list-error propagation, PCM widths/accounting, mastering gain and all tests alongside this PR's Gain early errors and callback safeguards. No open prerequisite is required.

#11352 independently proposes the same Gain-ramp correction already included here; apply that correction once and preserve the combined coverage whichever PR lands first. Custom Reverb #10599, experimental PCM retention #5548, and newer custom-module/channel/codec work remain separate. Their later compositions must preserve each topic's scope, state ownership and test registrations and be validated at their actual heads. The refreshed overlap report and independent reviews record current pinned references; no sibling branch was edited or combined-runtime result claimed.

Tested

Windows 11 Pro 10.0.26200; WinLibs MinGW-w64 GCC 15.2.0 r7; Ninja Release, BUILD_TESTING=ON, SPIRV-Tools enabled. Latest exact head: c40ba98945e845b950ecca7819c472942757ad30, main ancestor 9b520147fbeddff4357d6e8f92e5235a16342171.

• Reconfigured the existing isolated full-build tree and rebuilt both production NGS2 library targets plus all 22 focused test executables with cmake --build build --target <both libraries and NGS2 test targets> --parallel 2.
• ctest --test-dir build -R '^ngs2_' --output-on-failure --parallel 2: 30/30 passed, zero skips, 2.95 seconds. This includes all ten supported CustomSafety cases, Gain/Routing against both libraries, native/core PCM and native voice events.
• Retained literal/rational expectations for previous-level ramps, target replacement, channel/voice independence, pause/resume, grain changes, levels outside +/-32, float nonfinite classifications, finite integer output, invalid module positions/list termination, all supported channel layouts, 1-4 buffers, overwrite/aliasing, per-grain clearing and separate-buffer unity copying.
• python -B tools/check_conventions.py --base origin/main and exact-base whitespace checks pass. python -B core/libs/tests/test_public_headers.py --audit-only: 368 libraries/12,231 exports verified; this is a declaration audit, not a compiled-header run.
• Fresh independent GPT-6.1-Sol/xhigh macro and micro reviews at this exact head found no actionable candidate defect and confirmed this fulfills Dean's requested scope. They performed source/artifact checks; the exact-head runtime results above are parent-run.

The former unsupported-refusal tests and their historical mutation results are not retained-test evidence. No new title, GPU, original-module or PS5 run was performed for this follow-up. The earlier a43a610 production smoke reached a capped Dead Cells loading screen; it establishes no latest-head title result, audio fidelity, callback timing, GPU validation or performance. Full-suite/hosted runtime checks and unselected external compositions are not claimed.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

[Felipebmaitan]

Owner coordination for this standalone audio-routing change, currently a43a610bbf8a8c5b9e7499dd2631247a8597a551 on main e7e3e9c1984c293f06bff08428b8c89b7b932ac2. It has no prerequisite PR and remains draft. I own the new buffers, Gain changes and callback safeguards here; I have not edited any sibling branch.

@JJenkx (#5548), @JAB-dev (#9926/#10410), @kyon-eth (#10599): the overlap scan identifies shared test registrations, event/control propagation, the integer-output guard and custom-module setup/processing. Could you confirm ownership and landing order? Maintainers can choose an order if the later branch preserves every topic's behavior and tests. I will validate whichever composition is selected; none has been built or claimed green yet.

Two independent adversarial reviews and my source inspection confirmed these issues at the pinned counterpart heads:

• fix(libSceNgs2): clamp the mastering gain to +-32 and keep NaN, as the PS5 module does #10410 at 182e2a227f47286536d9852340badc8e088e5581: TestMasteringGainClampedLikePs5's chain helper creates a local PCM vector, submits its borrowed pointer, and returns before the render. The vector dies while the sampler retains it. The fixture needs storage alive through the final render/system destruction; fix(ngs2): retain PCM input when reporting decoded bytes #5548's first-decode copy cannot recover already freed input. Exact helper.
• feat(ngs2): implement the custom Reverb module #10599 at b24bea85547585f4350f1b59fc5a5a817ecbd13d: Reverb dispatch recognizes the eight-bit parameter-position family but extracts only five bits. 0x40001b20 aliases module zero. Exact dispatch. The published module reading gives 0x804A835A for an out-of-list position. A regression should also preserve state and stop a linked parameter list at that failure.

Could the respective owners take those repairs, or confirm who should own them, before we attempt shared edits? This is a static-source report, not a console measurement. The other audited heads were #5548 ddf93fbe14b4f036131e5c870a43874414414597, #9926 2375223a92690cfb798e4c9906dbc8eb5493e71b, #1384 70b4d2de4b4e9ed4c27259f4ef75f0e9f7b366e0, #10249 a4d01ff96ffaf96a9b7afa45f05cecaffbb0e4f9 and #10127 749f6cfa6586f6a515dcad24b60f20d2448423f2; none contains this branch's route/layout repair. That absence does not imply each standalone branch supports the new two-buffer reproducer.

On this branch, all four native/core route-mutation regressions fail at 2da1332bff0194443e91c9ab11a644cf6b4f58d8 and pass after the explicit refusal. The 27-test NGS2 suite passes. A further test-only tightening now distinguishes each safety refusal and adds nonfinite integer-output coverage; it will be pushed after validation and fresh review. No complete merged composition or PS5 fidelity is claimed.

[MarcZX10]

Reviewed the exact-head follow-up e50cbf349d4e0af8647627640df9c8c9ddf610b7 against a43a610bbf8a8c5b9e7499dd2631247a8597a551. This follow-up is test-only: callback safety cases now require the exact refusal and a single callback invocation, and the integer-output case checks finite positive/negative gain plus nonfinite rejection with output sentinels unchanged. I found no issue in these added assertions. I did not build or run them locally; the 27/27 NGS2 results in the description are author-reported. This is not a full review or approval of the composed change: the PR remains draft while overlap owners coordinate, and the body says no combined branch has been validated.

[Felipebmaitan]

Recommended landing queue, after refreshing current PR metadata and inspecting read-only source merge conflicts on 2026-10-25:

#9926 → #10127 → #10851 → repaired #10410 → repaired #10599. Defer draft #5548 until its own evidence gaps are resolved.

This replaces the open-ended landing-order question in my earlier coordination comment. It is my recommendation, not an agreement from other authors or a claim that any counterpart is ready to merge. This PR has no prerequisite and is being marked ready for standalone review; if an earlier item is unavailable, it can land independently. I accept ownership of rebasing #10851 after earlier changes and preserving their behavior/tests. No sibling branch was edited.

Why this order:

• fix(libSceNgs2): return INVALID_VOICE_ACCESS for PLAY on a voice that was never set up #9926 is the small, module-measured lifecycle fix with a reported title crash; establish its PLAY rejection, error propagation and all-zero-setup semantics first.
• feat(ngs2): play additional PCM sampler formats #10127 is already reviewed and its maintainer-requested test follow-up is addressed. Establish format-aware PCM strides/accounting before experimental retention. Its exact head source-merges cleanly with feat(ngs2): route custom submixer buffers through Gain #10851.
• feat(ngs2): route custom submixer buffers through Gain #10851 supplies the bounded buffers, Gain semantics and callback safeguards before the broader custom-Reverb integration. Its standalone final macro/micro reviews have no remaining actionable candidate finding; the existing exact-head 27/27 focused test result remains the runtime evidence.
• fix(libSceNgs2): clamp the mastering gain to +-32 and keep NaN, as the PS5 module does #10410 has separate mastering-gain semantics but duplicates the output safeguard. Its owner must first repair the freed-PCM test fixture. When rebasing, retain the mastering clamp/NaN behavior and feat(ngs2): route custom submixer buffers through Gain #10851's broader !isfinite integer refusal; do not restore the weaker NaN-only guard.
• feat(ngs2): implement the custom Reverb module #10599 should then adapt its per-module Reverb state, PLAY/setup reset and DSP to the landed buffer/revision guards. Repair the full-byte position validation and error/list-termination contract first; keep Reverb limited to its supported one-buffer layouts.
• fix(ngs2): retain PCM input when reporting decoded bytes #5548 remains an experimental draft for unmeasured decoder granularity/repeat/waveform-pointer timing. Its rebase must use format-aware retained-frame strides, count bytes once at the ownership transition, preserve route/layout protections and retain all tests. It should not delay this standalone PR.

Pinned inputs (main is still e7e3e9c1984c293f06bff08428b8c89b7b932ac2; no declared prerequisites):

PR	Exact inspected head	Source combination with #10851
#10851	e50cbf349d4e0af8647627640df9c8c9ddf610b7	Clean against current main; reviewed implementation unchanged
#9926	2375223a92690cfb798e4c9906dbc8eb5493e71b	Voice.cpp and native test list conflict; retain both Gain return/list termination and Ngs2ReturnCode event propagation
#10127	749f6cfa6586f6a515dcad24b60f20d2448423f2	Source merge clean; still requires combined runtime tests if composed
#10410	182e2a227f47286536d9852340badc8e088e5581	Render.cpp integer-output guard conflict; preserve the stronger guard and distinct mastering behavior
#10599	b24bea85547585f4350f1b59fc5a5a817ecbd13d	Custom.cpp, native test list and debt conflict; preserve both setup invalidation/revision and Reverb reset, dispatch and state
#5548	ddf93fbe14b4f036131e5c870a43874414414597	Both NGS2 test lists conflict; preserve every registration

I rechecked the two sibling defects at these unchanged heads: #10410 borrowed-PCM fixture, and #10599 position mask. @JAB-dev and @kyon-eth own those repairs unless they coordinate otherwise. @JJenkx: retention is not a prerequisite here. Owners/maintainers may choose a different queue; the later branch must preserve each topic and all registrations, then build both NGS2 libraries and run the union of their focused tests.

#1384 at 70b4d2de4b4e9ed4c27259f4ef75f0e9f7b366e0 remains a conflicting experimental draft, excluded from this queue; preserve current main's reference-based Reverb if revived. #10249 at a4d01ff96ffaf96a9b7afa45f05cecaffbb0e4f9 has no NGS2 source conflict with #10851 and no audio dependency; its merge check reports unrelated AGC/JobManager conflicts, and the maintainer has requested scope/regression work. Parser PRs #9148 (8accc665b121f428380c0545578bb8d2a94ebabc), #9149 (ea64374d3a35da14ee6259f7a8b9c6b7155fa43d) and #9324 (7644d6c2b4649f7f7c5b7b53e8faa71616632ab0) share only debt text with this patch and do not impose routing dependencies.

Validation for this documentation/readiness update: current heads/state/comments were refreshed; git merge-tree --write-tree --name-only --no-messages <headA> <headB> inspected the above source combinations without checking out or modifying another branch. No combined build/runtime result is claimed. No source, head SHA, test results, console evidence or prerequisite changed. No new tests were run because this update changes only PR metadata. AI-assisted: yes.

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
PR 5548	core/libs/prx/libSceNgs2.native/CMakeLists.txt L5 core/libs/prx/libSceNgs2/CMakeLists.txt L5-13
PR 9149	docs/dev/debt/libSceNgs2.native.md L49-50
PR 10599	core/libs/prx/libSceNgs2.native/CMakeLists.txt L5 core/libs/prx/libSceNgs2.native/src/Custom.cpp L95-97, L166, L245-250 docs/dev/debt/libSceNgs2.native.md L32-33, L35
PR 11347	core/libs/prx/libSceNgs2.native/src/Custom.cpp L166 docs/dev/debt/libSceNgs2.native.md L28-29, L32-35
PR 11352	core/libs/prx/libSceNgs2.native/src/Custom.cpp L199-204 docs/dev/debt/libSceNgs2.native.md L32-33
PR 11354	core/libs/prx/libSceNgs2.native/src/Ngs2Internal.hpp L149 docs/dev/debt/libSceNgs2.native.md L32-33
PR 11388	core/libs/prx/libSceNgs2.native/CMakeLists.txt L5

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 1384	core/libs/prx/libSceNgs2.native/CMakeLists.txt core/libs/prx/libSceNgs2.native/include/Ngs2Types.hpp core/libs/prx/libSceNgs2.native/src/Ngs2Internal.hpp core/libs/prx/libSceNgs2.native/src/Render.cpp core/libs/prx/libSceNgs2.native/src/Voice.cpp core/libs/prx/libSceNgs2/CMakeLists.txt
PR 10249	core/libs/prx/libSceNgs2.native/src/Custom.cpp
PR 11315	core/libs/prx/libSceNgs2.native/CMakeLists.txt core/libs/prx/libSceNgs2.native/include/Ngs2Types.hpp core/libs/prx/libSceNgs2.native/src/Custom.cpp core/libs/prx/libSceNgs2.native/src/Ngs2Internal.hpp core/libs/prx/libSceNgs2.native/src/Render.cpp core/libs/prx/libSceNgs2.native/src/Voice.cpp core/libs/prx/libSceNgs2.native/tests/Ngs2Custom.cpp docs/dev/debt/libSceNgs2.native.md

[oneandonlydean]

@Felipebmaitan The custom submixer buffers and the corrected Gain stage read well. One test change:

Several cases the console accepts are left as NotImplemented in the code, which is a fine placeholder, but the tests assert that they throw, with the exact messages:

• gain processing before the first parameter set, or after setup/stop;
• patch changes during source rendering;
• integer rendering of non-finite samples;
• voice setup changes during custom processing.

Each of those assertions would fail as soon as the case is implemented. Could you remove them (keep the NotImplemented in the code) and keep the rest of the Ngs2Custom, Ngs2CustomSafety and Ngs2Gain coverage? The branch also needs a merge of main (it currently conflicts).

[Felipebmaitan]

@oneandonlydean Addressed in c40ba98945e845b950ecca7819c472942757ad30.

• Removed the assertions requiring unimplemented Gain lifecycle, callback setup/route mutation and nonfinite integer rendering to throw, including the exact-message matching. The runtime guards and debt entries remain.
• Retained the rest of Ngs2Custom and Ngs2Gain, including finite integer output. Ngs2CustomSafety now checks supported custom/upstream signal callbacks, explicitly configured Gain, callback counts, port metadata and literal output. Its route changes happen before rendering; the block-end callback only counts invocations. All five supported cases run against each library.
• Merged main 9b520147fbeddff4357d6e8f92e5235a16342171 in 0c3f104067239abd5358b8d1992295f9da7b27ee. The resolution preserves main's PCM formats/accounting, PLAY/list errors, empty-setup behavior, mastering-gain/PCM-fixture repair, pan changes and every test registration alongside this PR's Gain errors and safeguards.

Reconfigured and rebuilt both production NGS2 variants plus all 22 focused executables on Windows/WinLibs GCC 15.2.0, with build concurrency 2. At exact head c40ba98945:

ctest --test-dir build -R '^ngs2_' --output-on-failure --parallel 2
30/30 passed, zero skips, 2.95 seconds

Conventions and whitespace checks pass. The declaration audit passes for 368 libraries/12,231 exports. No new title, GPU, PS5 or original-module result is claimed. Historical refusal-mutation results are not evidence for the revised tests.

Both fresh independent macro and micro reviews found no actionable candidate defect and confirmed the follow-up fulfills your request. Their source/artifact audits are separate from my exact-head build and 30/30 runtime execution above. This PR remains ready for standalone review; no sibling-owner agreement is a prerequisite.

Independent review pins and integration notes

Both reviewers use GPT-6.1-Sol/xhigh with independent isolated worktrees. They are read-only; the parent owns all edits and the exact-head build/test result above. Reviewed base: 9b520147fbeddff4357d6e8f92e5235a16342171; head: c40ba98945e845b950ecca7819c472942757ad30; prerequisites: none. The binary patch artifact SHA256 is fe4f9cbd57f716ff035b473636636736b6af5cbcff8b9e4ff86f9f7d0c56bd0a.

• Macro task: node:delegated-task:command%3Amcp%3Aaf0a8e9b-fd5a-4833-b3ae-e1558c384efd%3Adelegate-task%3Angs2-dean-macro-c40ba98945-r1
• Micro task: node:delegated-task:command%3Amcp%3Aaf0a8e9b-fd5a-4833-b3ae-e1558c384efd%3Adelegate-task%3Angs2-dean-micro-c40ba98945-r1

The previous queue's #9926, #10127 and #10410 are now merged into main. Newer #11352 at 66bca2376f64b8846397063c646f4f1c57673730 duplicates the Gain-ramp correction already present here: @edubox, either order should apply that correction once and preserve coverage; it is not a prerequisite of this PR. The refreshed overlap scan also covers #5548, #10599, #11315, #11347, #11354 and #11388. Existing conflicted #1384/#10249 comparisons remain manual integration work. No sibling checkout/branch was changed and no external composition is claimed tested.

The micro reviewer reverified two separate sibling parameter-index issues: #10599 at 54ac175ee309d0a7a517d7bd3524413e6ad8383c still extracts five bits for Reverb, and #11315 at 316f48500a7193fb89aae6d3d9b2c260df55b409 does so for immediate/deferred PitchShift. In each, position 0x20 aliases module zero and its positive fixture relies on the alias. @kyon-eth owns those branches: validate the full position before mutation/enqueueing, return 0x804A835A when invalid, preserve state and stop the linked parameter list, and use valid positions in positive fixtures. These are defined error-contract checks, separate from the unimplemented-path assertions removed here. Neither is included in or blocks standalone #10851.

The latest observed main 7704178631f3cb40e9d9c0efa168209f53cd27c3 source-merges cleanly with this head and changes no NGS2 files from the merged base. The exact runtime result above remains for c40ba98945, not a claimed build of that newer composition. Console initialization/numerical equivalence, callback timing, blocking-callback progress limits and broader custom chains remain the documented gaps.

AI-assisted: yes.

[oneandonlydean]

@Felipebmaitan Thanks, removing those assertions while keeping the guards is what was asked. We'll review the update and the main merge, then run CI on this head.