Skip to content

SonKu P7.1: measured direct-source XYZ and six-boundary geometry, isolated native research - #681

Closed
masarray wants to merge 7 commits into
feat/sonku-p7-measured-six-surface-20261114from
feat/sonku-p71-xyz-direct-and-fir-20261114
Closed

masarray wants to merge 7 commits into
feat/sonku-p7-measured-six-surface-20261114from
feat/sonku-p71-xyz-direct-and-fir-20261114

Conversation

@masarray

Copy link
Copy Markdown
Owner

Architecture / scope

Research branch stacked on P7 #679 → P6 #678 → P0–P5 #674. Production, GUI, VST3 catalog, native bridge, installer and .smix unchanged.

  • Real measured DIRECT elevation for a source placed at nonzero vertical angle (MIT KEMAR original 14-elevation tables), including central coherent MID correction; same authoritative native PCM path, no new FIR, no fake EQ/panner or proprietary dearVR coefficients.
  • Physical 3D image arrivals: radial distance remains 1.6–10 m and is projected into horizontal distance + source z in the same room. Existing four-wall reflection image sources gain the correct vertical distance/elevation and measured KEMAR angle; two extra floor/ceiling source images use causal path delay relative to the actual direct 3D path.
  • Fail-closed quiescent-only setResearchDirectElevation() exists ONLY when SONKU_P71_DIRECT_XYZ_RESEARCH=1 and the P7 six-plane profile is explicitly prepared; rejects invalid geometry and any edit after PCM starts. Not published as an unsupported UI/ACK or state version.
  • Baseline native P0–P7/0° rendering remains compiled identically; extra measurement/data/profile is research-only.
  • New 44.1/48 kHz tests for coordinate projection, mirror symmetry, 64/257 frame PCM parity, moving azimuth, stereo anti-phase and clipping, and research mutation/restart. Twelve original FRONT/REAR up/down/horizontal A/B/C WAVs at 48k/16bit; no subjective superiority claims.

P7.1 CPU/qualification remaining

  • Measure exact per-path CPU optimization versus P7, including Linux 1/2/4 insert p99 and PCM equivalence. No arbitrary callback-time auto switching.
  • Revalidate focused Windows JUCE P6 production-path opt-in (not P7.1 physical ASIO) and full scoped SonKu tests.
  • Owner listening matched SPL against dearVR, multi-insert ASIO/XRUN/clock with WIRE+VST3 and explicit UI/ACK/session design are still separate gates. Do not merge or ship without evidence.

@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: 9332029
Status: ✅  Deploy successful!
Preview URL: https://af2e2844.arsonmix.pages.dev
Branch Preview URL: https://feat-sonku-p72-native-xyz-ac.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

P7.1 engineering milestone CLOSED (research implementation) — 2026-11-14

Exact source SHA: 9332029e3345a4bab37a4dba0f06461a38724513 on draft PR #681, stacked on #679 → #678 → #674. This is NOT a shipped product, nor a dearVR parity certification.

Implemented and verified

  • Original measured MIT KEMAR HRIR for the direct binaural source at approved quiescent elevations, using true radial distance → source X/Y/Z projection. Measured L/R and coherent MID correction use one authoritative native PCM graph, no additional direct FIR. Physically consistent source-height geometry for four walls + floor/ceiling and causal path lengths.
  • Research control rejects non-finite/outside-room elevation and rejects mutation after any PCM, including after reset(). Only prepare() can re-open quiescent research admission. Absolutely no published new GUI, FX ID, native ACK, .smix version, WIRE, VST3 authority, installer or main change; P7.1 remains compiled out by default.
  • CPU hot path splits full warm ring FIR from history epoch warm-up to avoid stale-history conditional inside every 140-tap convolution. Every coefficient still slews and the full KEMAR 140 taps are retained; no FFT, HRTF truncation, callback heap/lock/time query or new audio graph.

CI facts from exact source SHA

  • Linux native source and all P0–P7 regressions: Action #43559215758 — SUCCESS. 44.1/48k, real 3D direct elevation, mirrored KEMAR, 64/257 chunk-parity, no post-publish edit, zero ghost on ROOM reactivation. Full 8-WAV old-P7 vs fast-P7.1 FIR PCM: max 1 LSB; 12 original 48k/16-bit FRONT/REAR source-horizontal/+25°/−25° WAVs; musical per-ear magnitude, ILD and interaural phase metrics. Download auditable native WAVs, JSON reports, CPU experiments (2-week CI retention).
  • Windows MSVC source-only: Action #43559215765 — SUCCESS, exact SHA verified, SonKuP71TrueXYZNativeWin + SonKuP6NativeDefaultWin: 2/2 Release CTests PASS. Minimal C++20 standalone CMake, no JUCE/WebView2/installer.
  • Previously closed P6 Windows JUCE adapter: Action #43554868473 — SUCCESS, 2/2 P6 real JUCE Release CTests with directional opt-in ON. This does not mean P7.1 experimental direct elevation is integrated into the JUCE plugin.
  • Interleaved FIR microbenchmark (48k/128, 4 inserts, 60° stereo spread, active room): 12 runs in ABBA×3 order, same synthesized source, steady warm history; median old/fast wall-time ratio 1.044× and median old/fast p99 ratio 1.242× on this Linux hosted runner. Previous non-interleaved observations varied in both directions; never infer Windows ASIO headroom or a deterministic universal 24% p99 gain.

Open product-release veto; move to separate qualification work

  1. Actual ASIO/WIRE/VST3 hardware and 1/2/4 insert 44.1/48k, 64/128/256 callback p95/p99/XRUN, device swap/restart; enforce explicit host admission, never naive callback-time auto-switch.
  2. Blind, matched-volume human headphone A/B against dearVR for front/rear, elevation, distance, real music tonal integrity, anti-phase, perceived externalization, head-locked stereo rescue.
  3. Reviewed real-time UI→native ACK/state/.smix 3D XYZ workflow (versioned/compatible), user-selectable licensed HRTFs, and licensing/installer third-party notices. Research-only method is not a user feature.
  4. Reconcile stacked draft PRs with fresh moving main and hardware sign-off before any merge/release.

Closure: P7.1 measured-XYZ research + qualified Linux/Windows source-level milestones are complete, and this ChatGPT development thread can be closed. Keep PR #681 and its ancestors draft, production P7/P7.1 OFF; open remaining release gates in separate tracking work.

Copy link
Copy Markdown
Owner Author

MILESTONE A — archived historical P6/P7 experimental SonKu stage (CLOSED UNMERGED). The original measured-KEMAR/XYZ study remains traceable and is not authorized as a product change. Listening-approved SonKu P9.9 was integrated to main through selective PR #733; older geometry research was never to be bulk-merged. Source branch feat/sonku-p71-xyz-direct-and-fir-20261114 and exact SHA 9332029e3345a4bab37a4dba0f06461a38724513; original PR target feat/sonku-p7-measured-six-surface-20261114. Registry: docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md on chore/milestone-a-pr-history-20261117. This closure does not delete source, alter checkpoint refs, merge code, retune audio, or change release. Reopen only for a separately approved verified feature.

@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