Skip to content

SonKu P8.4 Stage 2 (research): single-core musical HRTF fusion, binaural diagnostics and real WAV A/B - #704

Closed
masarray wants to merge 44 commits into
feat/sonku-p83-physical-speaker-mono-stereo-20261114from
research/sonku-p84-musical-clarity-anchor-20261115
Closed

masarray wants to merge 44 commits into
feat/sonku-p83-physical-speaker-mono-stereo-20261114from
research/sonku-p84-musical-clarity-anchor-20261115

Conversation

@masarray

@masarray masarray commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

Stage 2: music-first P8.4 native single-core candidate — SOURCE VERIFIED, HEADPHONE FAT NOT APPROVED

Why: User's physical Windows mixer headphone FAT rejected P8.3 as quiet, hollow and constrained versus the approved P8.2. PR #701 stays rejected. P8.2 is the non-negotiable musical baseline; there is no dearVR equivalence or externalization approval yet.

What is actually implemented

  • SonKuP84FusedMusicHrir.h and an opt-in single SonKuSpatialCore PCM graph. Precompose a 12% high-passed, 750Hz directional HRTF difference into the native left/right FIR target coefficients before a single native direct 2x2 convolution. No dual runtime DSP graph, invented LF480 matrix, makeup gain, extra audio path or filter allocation in callback.
  • Preserve P8.2 transparent stereo bass, SIDE, center music HRIR and room presence. Native early-reflection yaw follows fused HRTF yaw; stage still uses a head-locked music anchor and no actual head tracker.
  • P8.2 OFF and fusion enabled/disabled are isolated by a compile-time SONKU_P84_SINGLE_CORE_FUSION_RESEARCH flag default OFF plus pre-PCM-only research setter. No JUCE editor, session/bridge, public installer or release change.
  • sonku_p84_fused_single_core_test.cpp: 28 cases, 280 tone probes and 88 1/3-octave impulse-band measurements, 44.1/48k, seven program types and ROOM ON/OFF. Worst global RMS delta 1.264dB, largest band delta 2.204dB vs original P8.2; 64/128 PCM parity. sonku_p83_music_bed_byte_parity.cpp compiled both with/without the new P8.4 macro yields bit-identical P8.2 PCM across room and phase settings.
  • sonku_p84_spatial_impulse_evidence_test.cpp: real native front/rear and L/R direct impulse differences, ILD +/-16.6dB at 48k +/-90, IACC 0.39 and signed lag +/-35, nonzero world-yaw direct/late-room changes; not human externalization proof.
  • sonku_p84_fused_hosted_cpu_benchmark.cpp: 18 CPU cases at 44.1/48k 64/128/256, 1/2/4 inserts. On one Linux runner 48k/64/4, P8.2 median 744.2µs, P8.4 median 785.8µs (~+5.6%), P8.4 p99 833.16µs, zero sampled deadline misses. No physical Windows ASIO XRUN claims.
  • NEW tools/sonku_p84_offline_audition.cpp: real native, single-core per output streaming stereo WAV tool, compares P82_Golden against P84_FusedYaw0 and yaw +/-60, true unnormalized output WAV and CSV RMS/peak. PCM16/24/32 / float32 + DAW WAVEFORMATEXTENSIBLE, ROOM OFF/ON smoke checked. Source-only audition, not a Windows mixer build.

CI evidence

Explicit release and dependency veto

PR #704 currently STACKS on rejected #701. Do NOT merge this chain. This is a research sandbox. If next real-music/headphone tests accept the direction, extract only P8.4 fused measured transfer and necessary geometry into a clean P8.2-based branch; abandon P8.3 Physical/LF480 and rejected JUCE host. Future qualifying steps are headphone/source-specific EQ and licensed BRIR, coherent direct/early/late tracking, larger real-song/STFT tests, one JUCE realtime SPSC/ACK integration and physical blind front/back/elevation fatigue and ASIO XRUN testing. Until then, P8.2 golden is still the only owner-approved quality baseline.

Engineering notes: docs/SONKU_P84_CLARITY_FIRST_ARCHITECTURE_20261115.md.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 14, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: 641b4ea
Status: ✅  Deploy successful!
Preview URL: https://a519c2bd.arsonmix.pages.dev
Branch Preview URL: https://research-sonku-p85-world-loc.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

P8.4 native music-fidelity research checkpoint (15 Nov 2026 WIB)

Real C++ native source test PASS, not JUCE/ASIO headphone qualification. The current isolated research branch HEAD is fe4f1c4a47e4bfe5a1ed1b284811d916a90e649c (documentation-only update after source f531324cb3f9a93f2f8ed066f0f9cd1914af9883).

Verified source run https://github.com/masarray/arsonmix/actions/runs/43640451156:

  • Golden P8.2 exact PCM when influence=0; 12% directional contrast anchored to actual P8.2 DSP; measured P8.3 physical reference used only in the OFFLINE comparison.
  • 44.1/48k × ROOM OFF/ON × 7 input programs (mono, antiphase, L, R, partially correlated, transients, fullband impulse) = 28 cases.
  • 10 per-program tone frequencies = 280 spectral metrics; 1/3-octave broadband impulse sweep = 88 additional band metrics, with 5 sampled frequencies summed per band.
  • Worst impulse spectral 1/3-octave color delta −2.182dB at ~3.63kHz (48k ROOM OFF), max program RMS deviation 1.267dB, both within strict existing bounds (2.5dB frequency / 1.5dB program); all finite, 64/128 PCM parity and head-yaw differential PASS.
  • Failed 18% candidate remains documented: 5.387kHz −3.208dB vs P8.2. Reduced influence to 12% from empirical evidence, without lowering gate.

Hard limitations and next gate: This is a 2-core offline prototype only; never deploy that architecture directly to live mixer. Passing PCM timbre constraints is NOT proof of externalization—head-locked P8.2 anchor could still sound inside the head. Before any Windows portable: fuse a coherent single-native graph, test measured source-specific BRIR and head tracking, CPU 1/2/4 inserts, wideband ILD/ITD/IACC/DRR, head-turn and blind headphone perceptual FAT. The P8.3 physical Windows artifact and PR #701 are STILL REJECTED by user. Keep P8.2 approved baseline intact.

* fix(gui): unify Preview/native console geometry, fader calibration and subdued gain styling

Render EXT rows in the single approved Pages source with browser-only hardware disabled, share routing heights and chrome across WebView2/Pages, reuse native tapered travel in both contexts, and harmonize GAIN metal/LED/LCD to the fader color system. Preserve native ACK/readback, WIRE DSP, PNG knob, zero-detent and snapshot behavior. Add parity/Chromium assertions.

* test(gui): move native visual assertions to shared Pages and correct unity pointer position

Keep hardware ACK tests intact; fader taper and EXT CSS now belong to canonical Pages, so regression guards must inspect the shared source. Update browser tactile unity test to the same 77% scale used by EXE.

* test(gui): expect cyan-steel focus halo instead of legacy green GAIN glow

* test(gui): align legacy V7 utility smoke with approved PFL semantic button

Legacy helper upgrades SOLO SVG, but approved PFL uses data-hw-kind=pfl / data-action=cue and a PNG in assembled Pages. Check the current semantic fallback instead of the removed cue-key SVG.

* test(gui): align V6 hardware-key smoke with current approved PFL semantics

Keep historical MUTE and SOLO SVG/CSS coverage; replace retired two-SVG cue assumptions with approved semantic PFL checks. Assembled /preview/ PNG artwork stays authoritative. No GUI or PCM changes.

* test(gui): preserve MUTE material contract with current semantic PFL fallback

Historical raw Pages uses one SOLO SVG. The real PFL button is PNG-owned in assembled /preview/, not a second SVG. Keep independent mute/solo/PFL action checks.

* test(gui): calibrate M1 ruler against shared Preview/EXE physical fader taper

* fix(gui): stop calibrated fader ruler numerals overlapping at 1366px

* fix(gui): preserve owner-frozen fader hardware while clarifying tapered dB labels

* fix(gui): preserve tabular numerals while spacing tapered fader labels

* fix(gui): fit shared four-row IN/OUT/EXT routing and preserve Master two-key QA
@masarray masarray changed the title SonKu P8.4 (research): timbre-constrained native directional contrast anchored to P8.2 SonKu P8.4 Stage 2 (research): single-core musical HRTF fusion, binaural diagnostics and real WAV A/B Nov 14, 2026

Copy link
Copy Markdown
Owner Author

P8.4 Stage 2 evidence — actual one-native-PCM-core (15 Nov 2026)

Draft PR #704 at source commit 1d9f2d702a2decbd0f19e1ff230b41e0f5de4de2, not merged to main. Owner's bad P8.3 physical FAT remains a hard rejection.

Verified source CI:

  • https://github.com/masarray/arsonmix/actions/runs/43644595928 SUCCESS with a single SonKuSpatialCore fused measured 2x2 HRTF (12% high-passed directional FIR difference), preserving P8.2 original stereo bass/SIDE. Never instantiate two PCM engines in live JUCE.
  • Native P8.2 compile ON/OFF byte-exact (44.1/48k, ROOM OFF/ON mono/antiphase).
  • 28 × music-like/impulse source cases, 280 tone probes, 88 1/3-octave IR probes; worst RMS relative difference 1.264dB, worst IR band 2.204dB vs P8.2; 64/128 parity.
  • Directional impulse front/back, left/right, native 48k ILD -16.602/+16.602dB at ±90, IACC=0.39053 and opposite ±35-sample lag, plus nonzero direct/room tail head yaw. Measured direct/tail window ratio ~19.26dB is not standardized real-room DRR or headphone proof.
  • Hosted CPU 48k/64/4 inserts P8.2 p50 744.20 µs vs P8.4 p50 785.80 µs, P8.4 p99 833.16 µs, 0 deadline misses on measured Linux blocks, not Windows ASIO XRUN evidence.
  • New tools/sonku_p84_offline_audition.cpp renders user's own WAV to four independently rendered native P8.2/P8.4 + simulated yaw0/±60 WAVs and true RMS/peak CSV. GitHub CI checks ROOM OFF/ON and float32 WAVE fact output.

Containment: PR #704 is stacked on rejected P8.3 PR #701 and MUST NOT be merged through that dependency. If further measured BRIR and headphone-listening evaluation approves the concept, extract only P8.4 HRIR fusion onto a clean approved P8.2 branch with host v5/SPSC/ACK and verified world-fixed BRIR. Until then no P8.4 Windows portable, GUI control, installer or parity claim with dearVR. Owner-approved P8.2 is unchanged; no proprietary dearVR algorithm/data copied.

…ole (#705)

* test(gui): early real route overflow diagnosis for shared Preview/EXE

* fix(gui): fit shared IN/OUT/EXT rows in all console densities

* test(gui): align legacy PAN detent smoke with approved V61 hysteresis

Copy link
Copy Markdown
Owner Author

P8.4 native Windows FAT checkpoint (2026-11-15 WIB)

First full native Windows build run https://github.com/masarray/arsonmix/actions/runs/43652113706 compiled the actual ARMixConsole executable successfully and 9/10 scoped CTests passed, but the new v5 JUCE test failed at a test-fixture-only assertion (tests/sonku_p84_native_juce_v5_test.cpp:72). Root cause: the fixture asserted the size of a JUCE MemoryBlock before MemoryOutputStream's destructor flushed its final byte count. The v5 real audio output, native PCM difference, audio-applied ACK, device re-prepare, initial v5 write/read all passed earlier in the same test before the failing fixture. No artifact was uploaded from this failed run.

Fixed fixture by explicitly closing the MemoryOutputStream scope before validating v1-v4 sizes (commit c16666ac3ffa6247ac00a9673fed3ee150504f43). Hardening: old v4 modes 1/2 downgrade to golden P8.2 and all old LF480 flags are sanitized, even for old v4 mode 0. Native v5 decoder and restored session now reject LF480 flags. Added targeted migration test (commits 6beb2918fdb98105c46f8c50393d259a3aff7e79, 47d168ab804320b38be8ef40be78d9971e327502, adcb4f003689e809a0ffb47c701fbf8c5db2aaa2).

Reordered .github/workflows/sonku-p84-windows-field.yml to build/test JUCE v5 state/audio ACK first, before allocating full Windows GUI compilation, and added real live P8.4 renderMode3 source preflight. Latest candidate commit 641b4eaec653b2ecacc997200356a53aa98c4339, Windows run https://github.com/masarray/arsonmix/actions/runs/43653357348. Keep draft, no public release or dearVR parity claim until verified portable and physical headphone FAT.

Copy link
Copy Markdown
Owner Author

✅ P8.4 actual Windows mixer HEADPHONE FAT candidate — verified and downloadable

Exact branch head 641b4eaec653b2ecacc997200356a53aa98c4339 built and verified on GitHub Actions run #43653357348 (completed SUCCESS). Windows actual full JUCE application ARMixConsole_artefacts/Release/AR MIX Studio.exe compiled (15,055,360 bytes), 10/10 scoped native Windows CTests PASS, and verified full packaged GUI including embedded approved skins, WebView2, state/provenance integrity, MIT KEMAR notice and FAT instructions. Real native v5 JUCE callback test (SPSC queued→audio-applied ACK/readback; 12ms mode fades; session v1-v5 and old P8.3 physical safe migration; A/B different native PCM; device reprepare) PASS both as dedicated preflight and final CTest. Source P8.4 gate, world-yaw/room evidence and music fidelity passed on Linux. No physical ASIO headphone listening in CI.

Download Actions → Artifacts → ArSonMix-SonKu-P84-Native-Mixer-Windows-x64-43653357348, artifact ID 14137561665 (15,118,216 bytes). GitHub Actions upload ZIP SHA256 e025ac5ac6c3c92e032bc9e8c1151279e29fc775611fd2ecada0d3effa321b16. Portable internal verification also passed source fingerprint.

Owner FAT instructions: extract independently from P8.2 good baseline; run AR MIX Studio.exe, select actual stereo hardware ASIO/WASAPI, route stereo audio to a mixer channel with native 8D-SonKu insert → MAIN → headphones (no parallel dry path; master mono OFF). In SonKu native editor switch MUSIC BED · P8.2 GOLDEN versus SPATIAL · P8.4 MUSIC FIRST; first compare ROOM OFF, STATIC FRONT azimuth 0°, spread 60°, yaw 0°. Then test ROOM ON, manual simulated yaw +60°/-60°. Keep trim/gain/sound source constant. Explicitly evaluate vocal body, bass weight, stereo phase/width, total loudness, front/back externalization versus the previous P8.2 and dearVR. Do NOT confuse yaw slider with actual head tracking; this is an unqualified research candidate.

Safety: P8.3 Physical/LF480 rejected by owner's prior real headphone FAT is NOT offered in v5 UI/control path; old v4 mode1/2 sessions migrate to P8.2. P8.2 remains golden and main/public release was not changed. PR #704 stays DRAFT; do not merge while still based on rejected P8.3 research chain; future accepted integration requires clean P8.2-based extraction and owner physical headphone feedback. No dearVR parity claim.

Copy link
Copy Markdown
Owner Author

P8.4 Windows x64 FAT — verified artifact for owner headphone listening (2026-11-15 WIB)

Verified existing full native Windows x64 executable + user-downloadable GitHub Actions artifact at frozen commit 641b4eaec653b2ecacc997200356a53aa98c4339:

  • GitHub run: https://github.com/masarray/arsonmix/actions/runs/43653357348
  • Direct GitHub Actions artifact page: https://github.com/masarray/arsonmix/actions/runs/43653357348/artifacts/14137561665
  • Artifact name: ArSonMix-SonKu-P84-Native-Mixer-Windows-x64-43653357348, artifact ID 14137561665, ZIP ~15.1 MB, not expired at verification, GitHub expiration 2026-12-05 23:38 UTC.
  • Both workflow jobs PASS: Linux strict P8.4 audio preflight and actual full MSVC/JUCE Windows 2022 portable build. Windows 10/10 scoped native CTests PASS, including v5 mode3 audio-applied ACK and old state restore; ZIP/PE, exact Pages embedded GUI bundle, source SHA, 5-schema provenance and MIT KEMAR notice verified.
  • The FX user-facing picker in the built-in 8D-SonKu editor presents MUSIC BED · P8.2 GOLDEN and SPATIAL · P8.4 MUSIC FIRST when native schema v5 capability is admitted, no P8.3 Physical options exposed in this v5 UI.
  • Do not call this production, ready for release, user headphone approved, or equivalent to dearVR. No physical ASIO XRUN results, room realism and real head tracking are still unqualified. This branch is stacked on rejected P8.3 work and is not the latest main/preview GUI.
  • FAT protocol: unzip into a separate test directory; retain approved P8.2 install/session, compare a familiar stereo track at identical outputTrim/MIX/room/headphone volume in the SAME v5 portable, switch modes P8.2 Golden -> P8.4 Music First. Start STATIC (not ORBIT), head yaw 0°, spread 60°, ROOM OFF, then ROOM ON, 44.1/48k. Avoid normalizing/manually increasing P8.4 gain; note bass kick, vocal body, head/overhead sensations, front/back source plausibility, phase oddities and dropouts. Stop and revert to P8.2 if objectionable. No public merge/release without actual listening approval.
  • This is an existing verified build; no claim that any newer commit or main GUI was built into artifact.

This user FAT link and release veto are preserved here for reproducibility.

Copy link
Copy Markdown
Owner Author

MILESTONE A: CLOSED UNMERGED as historical P8.x SonKu research. This opt-in/private experiment predates the approved SonKu P9.9 source selectively integrated in #733; no permission exists to bulk-replay the earlier research ancestry into production. Original source branch research/sonku-p84-musical-clarity-anchor-20261115, immutable original SHA 641b4eaec653b2ecacc997200356a53aa98c4339, original target feat/sonku-p83-physical-speaker-mono-stereo-20261114. Head pins and rationale are recorded in docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md on chore/milestone-a-pr-history-20261117. No code merged, no branch/checkpoint deleted or changed, no audio retuned, no release republished. Reopen only with explicit new scope and approval.

@masarray masarray closed this Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant