Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
44 commits
Select commit Hold shift + click to select a range
4be9634
feat(sonku): add bounded phase-aware stereo music-integrity stage
masarray Nov 13, 2026
d40ce26
fix(sonku-dsp): retain anti-phase stereo music while preserving measu…
masarray Nov 13, 2026
f19940e
test(sonku): prove exact centered HRTF, anti-phase stereo energy and …
masarray Nov 13, 2026
5e14e38
test(sonku): publish reproducible lossless music-integrity headphone …
masarray Nov 13, 2026
a432d8d
fix(sonku): render real line breaks in offline headphone A/B listenin…
masarray Nov 13, 2026
119d37d
ci(sonku): support fast three-track music integrity headphone AB with…
masarray Nov 13, 2026
2d3715b
fix(sonku): correct literal WAV listening-guide newline escapes
masarray Nov 13, 2026
881f04c
docs(sonku): root-cause cancellation, public dearVR research and mult…
masarray Nov 13, 2026
a5f3ea8
ci(sonku): keep audio-only phase work in fast native WAV and CPU lane…
masarray Nov 13, 2026
f7cd174
test(sonku): honor existing 8ms wet/dry smoothing before asserting bi…
masarray Nov 13, 2026
9bf0f7c
feat(sonku-audio): detect coherent stereo mid separately from anti-ph…
masarray Nov 13, 2026
bed18ba
feat(sonku-dsp): phase-coherent Mid HRTF with zero extra per-sample FIR
masarray Nov 13, 2026
c105273
test(sonku): measure native coherent-mid 2 kHz HRIR vs previous rear …
masarray Nov 13, 2026
67ecb1b
test(sonku): render five matched native WAVs against immutable alpha.…
masarray Nov 13, 2026
59e4a27
test(sonku): validate pinned alpha.23 vs candidate PCM in five matche…
masarray Nov 13, 2026
fe1fa6c
ci(sonku): qualify spectral and anti-phase PCM against pinned alpha.2…
masarray Nov 13, 2026
27b384d
docs(sonku): record 7.2/11.6dB coherent comb mechanism and zero-extra…
masarray Nov 13, 2026
a1a8af0
docs(sonku): pin objective 7.19/11.64dB and anti-phase RMS from nativ…
masarray Nov 13, 2026
870e8b0
test(sonku): protect ROOM ON anti-phase source at 44.1/48k across arb…
masarray Nov 13, 2026
d4964e4
test(sonku): 380-point 44.1/48k KEMAR music spectral fidelity audit
masarray Nov 13, 2026
b0d32b2
test(sonku): normalize spectral dB against exact center magnitude eve…
masarray Nov 13, 2026
83e7836
ci(sonku): add fast 380-point spectral music guard to exact-source WA…
masarray Nov 13, 2026
9dd16d3
docs(sonku): record 380-point spectral integrity coverage and explici…
masarray Nov 13, 2026
0807b92
feat(sonku): protect mixed-polarity frequency bands in native side guard
masarray Nov 13, 2026
93e749b
test(sonku): native A/B WAV proves masked treble rescue without bass …
masarray Nov 13, 2026
ba11faf
fix(sonku): suppress multiband protection tail on exact digital silence
masarray Nov 13, 2026
3ee29c2
test(sonku): audit treble interaural phase, level and accurate headph…
masarray Nov 13, 2026
301ef48
fix(sonku): write actual line breaks in six-pair headphone listening …
masarray Nov 13, 2026
7fd6446
perf(sonku): measure native callback p50/p95/p99 for 64–512 frame buf…
masarray Nov 13, 2026
a1e6a12
fix(sonku): terminate per-buffer CPU percentile log lines correctly
masarray Nov 13, 2026
4992ee3
feat(sonku): prototype directional measured-KEMAR early reflections i…
masarray Nov 13, 2026
11b6d2e
test(sonku): P2 directional room same-source WAV A/B, objective evide…
masarray Nov 13, 2026
0e52b07
fix(sonku): ensure P2 headphone A/B guide uses real line breaks
masarray Nov 13, 2026
8f7a05c
fix(sonku): keep immutable alpha.23 baseline build independent of P2-…
masarray Nov 13, 2026
56dcd1e
perf(sonku): linearize directional reflection FIR reads with mirrored…
masarray Nov 13, 2026
af4e402
test(sonku): prove P2 moving parity and near/far; reject unproven rin…
masarray Nov 13, 2026
395e2fb
test(sonku): quantify P2 binaural harmonic coloration and tonality re…
masarray Nov 13, 2026
8761f78
feat(sonku): P3 zero-spread directional reflection fusion with exact …
masarray Nov 14, 2026
68d944a
test(sonku): preserve physically correct anti-phase null in P3 wet-on…
masarray Nov 14, 2026
96763f0
test(sonku): verify P3 native 140-tap reference/fused WAV phase and h…
masarray Nov 14, 2026
77d1fc9
perf(sonku): P4 opt-in contiguous full-KEMAR reflected FIR, cross-wra…
masarray Nov 14, 2026
333f5a6
test(sonku): certify P4 original independent-stereo native PCM byte i…
masarray Nov 14, 2026
c30e538
test(sonku): quantify P4 16-bit rounding, sample RMS and 5-band stere…
masarray Nov 14, 2026
5ae4d88
fix(sonku): P5 reject stale directional HRIR ring samples on ROOM rea…
masarray Nov 14, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
78 changes: 77 additions & 1 deletion .github/workflows/ui-smoke.yml
Original file line number Diff line number Diff line change
Expand Up @@ -113,6 +113,8 @@ jobs:
node --check site/preview-v2/assets/mixer-native-sonku.js
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_spatial_core_test.cpp -o /tmp/armix-sonku-core-smoke
/tmp/armix-sonku-core-smoke
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_multiband_phase_test.cpp -o /tmp/armix-sonku-band-smoke
/tmp/armix-sonku-band-smoke
python tests/native_peq_assembly_test.py
python tests/native_v60_fader_assembly_contract.py
python tests/approved_operator_ux_smoke.py
Expand Down Expand Up @@ -198,7 +200,7 @@ jobs:
needs: quick
# Both 8D-SonKu and SOF native-owner PRs share ONE actual Chromium runner.
# Preserve fast-only checks for unrelated GUI and engine branches.
if: ${{ (github.event_name == 'pull_request' && (startsWith(github.head_ref, 'feat/8d-sonku-') || startsWith(github.head_ref, 'feat/sonku-') || startsWith(github.head_ref, 'test/sof-native-') || startsWith(github.head_ref, 'integration/sof-sonku-main-') || startsWith(github.head_ref, 'fix/native-wire-inspector-') || (startsWith(github.head_ref, 'integration/sonku-m4-wire-') || startsWith(github.head_ref, 'integration/sonku-m4-guitar-picker-') || startsWith(github.head_ref, 'fix/sonku-')))) || (github.event_name == 'workflow_dispatch' && startsWith(github.ref, 'refs/heads/feat/8d-sonku-') && inputs.expected_sha != '') }}
if: ${{ (github.event_name == 'pull_request' && (startsWith(github.head_ref, 'feat/8d-sonku-') || (startsWith(github.head_ref, 'feat/sonku-') && !startsWith(github.head_ref, 'feat/sonku-music-integrity-')) || startsWith(github.head_ref, 'test/sof-native-') || startsWith(github.head_ref, 'integration/sof-sonku-main-') || startsWith(github.head_ref, 'fix/native-wire-inspector-') || (startsWith(github.head_ref, 'integration/sonku-m4-wire-') || startsWith(github.head_ref, 'integration/sonku-m4-guitar-picker-') || startsWith(github.head_ref, 'fix/sonku-')))) || (github.event_name == 'workflow_dispatch' && startsWith(github.ref, 'refs/heads/feat/8d-sonku-') && inputs.expected_sha != '') }}
runs-on: ubuntu-latest
timeout-minutes: 12
steps:
Expand Down Expand Up @@ -276,6 +278,80 @@ jobs:
if-no-files-found: ignore
retention-days: 7

sonku-music-integrity-audio:
name: SonKu music integrity native WAV/CPU (no Chromium, no Windows packaging)
needs: quick
if: ${{ github.event_name == 'pull_request' && startsWith(github.head_ref, 'feat/sonku-music-integrity-') }}
runs-on: ubuntu-latest
timeout-minutes: 9
steps:
- uses: actions/checkout@v4
# Public alpha.23 is immutable baseline; fetch ONLY SonKu engine
# headers for matched native C++ reference rendering (no JUCE/SDK).
- name: Checkout immutable public alpha.23 SonKu baseline headers
uses: actions/checkout@v4
with:
ref: 4a6e47e2905a4aa995f879045a907abee8b1aae5
path: .baseline-sonku
persist-credentials: false
sparse-checkout: |
engine/builtin_fx/sonku
sparse-checkout-cone-mode: true
- name: Render and numerically compare six native PCM music A/B tracks
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tools/sonku_offline_listening.cpp -o /tmp/sonku-integrity-audio
/tmp/sonku-integrity-audio sonku-integrity-evidence --music-integrity-only
g++ -std=c++20 -O2 -Wall -Wextra -pedantic -DSONKU_AB_BASELINE tools/sonku_offline_listening.cpp -o /tmp/sonku-integrity-baseline
/tmp/sonku-integrity-baseline sonku-integrity-baseline --music-integrity-only
python tests/sonku_music_integrity_comparison.py sonku-integrity-baseline sonku-integrity-evidence
- name: P2 measured-KEMAR early reflection source regression
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_directional_reflections_test.cpp -o /tmp/sonku-p2-er
/tmp/sonku-p2-er
- name: P3 directional zero-spread fusion PCM and spread-transition regression
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_p3_spread_zero_fusion_test.cpp -o /tmp/sonku-p3-fusion-test
/tmp/sonku-p3-fusion-test
- name: P4 ring-wrap/source-motion full-HRIR equivalence regression
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_p4_contiguous_reflections_test.cpp -o /tmp/sonku-p4-regression
/tmp/sonku-p4-regression
- name: P5 ROOM OFF ON fresh-history directional FIR protection
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_p5_room_reactivation_test.cpp -o /tmp/sonku-p5-room
/tmp/sonku-p5-room
- name: P2 original voice-like headphone A/B and numeric PCM comparison
run: |
/tmp/sonku-integrity-audio sonku-integrity-evidence --directional-er-only
python tests/sonku_directional_reflections_audio_audit.py sonku-integrity-evidence
- name: P3 exact-source FRONT/REAR native P2/P3 PCM A/B
run: |
/tmp/sonku-integrity-audio sonku-integrity-evidence --p3-fusion-only
python tests/sonku_p3_fusion_audio_audit.py sonku-integrity-evidence
- name: P4 original independent stereo 60/120 FRONT/REAR PCM identity
run: |
/tmp/sonku-integrity-audio sonku-integrity-evidence --p4-contiguous-only
python tests/sonku_p4_contiguous_audio_audit.py sonku-integrity-evidence
- name: Broad KEMAR frequency-response integrity audit (44.1k/48k)
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tests/sonku_hrir_spectral_audit.cpp -o /tmp/sonku-hrir-audit
/tmp/sonku-hrir-audit > sonku-integrity-evidence/KEMAR-SPECTRAL-AUDIT.csv
test "$(($(wc -l < sonku-integrity-evidence/KEMAR-SPECTRAL-AUDIT.csv)-1))" -eq 380
- name: Same-runner informational SonKu native realtime CPU check
run: |
g++ -std=c++20 -O2 -Wall -Wextra -pedantic tools/sonku_rt_benchmark.cpp -o /tmp/sonku-integrity-cpu
/tmp/sonku-integrity-cpu
- name: Upload original lossless music-integrity proof
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: ArSonMix-SonKu-Phase-Integrity-${{ github.run_id }}
path: |
sonku-integrity-evidence/*
sonku-integrity-baseline/*
if-no-files-found: ignore
retention-days: 14

vocal-throw-preview:
name: Vocal Delay native editor desktop/mobile Chromium FAT
needs: quick
Expand Down
78 changes: 78 additions & 0 deletions docs/SONKU_ADR_P1_1_MULTIBAND_PHASE_INTEGRITY_20261114.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,78 @@
# SonKu ADR P1.1 — mixed-polarity frequency-local integrity

2026-11-14 · experimental draft PR #674 · no public release.

## Source-proven failure and decision
The single JUCE native PCM `SonKuSpatialCore` maps L/R into two
measured KEMAR HRIR sources, summed at each ear. At zero source
spread, L=-R in *any frequency band* cancels in both ears. The
existing P0 full-band correlation guard recovers only globally
anti-correlated input. A stronger centered 170/220Hz bass can
hide an inverted 3200Hz instrument, leaving global correlation
positive and the high-frequency signal destructively lost.

Inside the **existing** callback-owned `SonKuStereoIntegrity`,
split each input with prepared 260Hz and 2600Hz one-pole filters:
`low=LP260(x)`, `mid=LP2600(x)-LP260(x)`,
`high=x-LP2600(x)`. These three bands sum to the original
input sample for sample (no linear-phase FIR, lookahead or new
graph). Each has its own 20ms correlation evidence, -0.78/-0.55
hysteresis and 15ms de-click blend. Preserve only locally
inverted stereo difference; the existing broadband gate remains
authoritative when full-band L=-R, to avoid double rescue.
The P0 positive coherent MID HRIR correction remains unchanged.

## ABI/performance/risk
Only small fixed POD state and per-sample add/multiply operations.
No allocations, locks, file I/O, network, runtime coefficient
design, FFT, extra convolution or GUIs. The original source
side-bed bound of 0.65×mix×trim, `.smix` v1/v2,
native ACK/readback, routing, FX identity and JUCE graph remain.

A post-HRTF stereo side bed is partly **head-locked**, not
physically externalized. Band crossover changes can affect
timbre/image stability. A passed unit test is only proof of
synthetic signal preservation, NOT externalization or dearVR
equivalence. Reject the experiment if subjective A/B shows
unacceptable artifacts. Subsequent milestones require directional
BRIR, personalizable licensed HRTF, distance DRR and real ASIO
CPU/headphone evidence.

## References
- Official dearVR PRO 2 manual (documented room/distance/Clarity
product behaviors, not proprietary code/coefficients):
https://files.plugin-alliance.com/products/dearvr-pro-2/dearvr-pro-2_manual.pdf
- Sennheiser official timbre-vs-localization discussion:
https://help.sennheiser.com/hc/en-us/articles/27424692616594-How-to-produce-natural-immersive-music-with-the-CLARITY-modul-of-dearVR-PRO2
- SOFA/AES69 (future data format): https://www.sofacoustics.org/
- 2026 BRIR/early reflection externalization evidence:
https://pmc.ncbi.nlm.nih.gov/articles/PMC13451706/
- ITU-R BS.1116-3 listening method:
https://www.itu.int/rec/R-REC-BS.1116-3-201502-I/en

SonKu's frequency-band decision is original experimental
filter/correlation algebra, not a claim about dearVR internals.

## Verified first-pass GitHub Actions evidence

Following exact digital-silence repair, [Actions #43524526143](https://github.com/masarray/arsonmix/actions/runs/43524526143)
passes native 44.1/48k source regression, six identical-input baseline/candidate
WAV pairs, 380 KEMAR spectral checks and runner-local CPU benchmark.
The sixth 3200Hz note recovered above an objective amplitude gate of
0.006 while the coherent 220Hz bass changed by less than 0.25dB.
This is a machine test threshold, not a perceptual listening rating.
The run's artifact is [#14083314097](https://github.com/masarray/arsonmix/actions/runs/43524526143/artifacts/14083314097).

## Informational block callback distribution (same native core)

The scoped `tools/sonku_rt_benchmark.cpp` collects external
`steady_clock` timing around each exact DSP `process` call, sorts
per-callback timings and reports p50/p95/p99 in microseconds on
64/128/256/512-frame blocks at 44.1/48kHz, for HRTF-only, early-room
and late-room modes. Also reports p99 divided by the theoretical
buffer duration. All benchmark allocations are **outside** the DSP
callback; no new runtime telemetry or DSP branch is added.
Hosted CI percentiles are informational and may include scheduler
jitter; there is deliberately no unstable hard CPU pass threshold.
Real Windows ASIO field p99, XRUN count and source+FX+WIRE integration
remain release blockers.
102 changes: 102 additions & 0 deletions docs/SONKU_ADR_P2_DIRECTIONAL_REFLECTIONS_20261114.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,102 @@
# SonKu ADR P2 — Directional measured-HRIR image-source reflections
2026-11-14 · experimental on draft PR #674 · NOT public/released

## Verified source-level gap
Existing SonKuEarlyReflections uses measured wall image-source paths
but ear output is a delayed filtered mono sample with synthetic
0.31ms maximum ITD and 13% ILD proxy. It omits the pinna/FRONT-REAR
spectral signatures of each reflected arrival. The existing KEMAR
bank provides actual 0° elevation measured horizontal HRTF.

## Independent design
Retain exact current realtime default. Add OFFLINE-only
setDirectionalEarlyReflectionsHrir(true) forwarded to the original
room instance (not a parallel audio graph). Active changes are rejected.
For each of 2 stereo source emitters x 4 first-order image paths,
compute image-source arrival azimuth atan2(x,y) every 64 samples.
Use horizontal KEMAR 5° HRIR interpolation with ear mirroring and
existing wall/distance pressure law; apply one COMMON path delay
and measured left/right KEMAR 140-tap FIRs. The measured HRIR
already includes ITD/ILD: no additional fake ear delay/gain.
Smooth moving filter targets over 8ms, existing delay/gain slew
remains; fixed state and history, no allocation, I/O or locks in
callback. Baseline renderer path stays structurally unchanged.
No new GUI, .smix, VST3 catalog, routing or JUCE ACK protocol.

## Scientific and commercial provenance
dearVR PRO 2 *official user manual*, §§2.1 and 3.2:
https://files.plugin-alliance.com/products/dearvr-pro-2/dearvr-pro-2_manual.pdf
describes first reflections and 6 surfaces; this is NOT a claim
about unpublished dearVR DSP. Steam Audio documentation:
https://partner.steamgames.com/doc/features/steam_audio
describes directional HRTF and reflected geometry, and associated
CPU costs. Meta spatial audio guide:
https://developers.meta.com/vr/design/spatial_audio/
explains spectral pinna cues in FRONT/REAR. SOFA/AES69:
https://www.sofacoustics.org/ for future multi-elevation BRIR
profile interchange. Existing SonKu KEMAR license remains under
docs/THIRD_PARTY_8D_SONKU_HRTF.md. No proprietary assets copied.

## Limitations and gates
This is a physically grounded HORIZONTAL early reflection pilot,
not full externalization, not 3D BRIR, not elevation or head tracking.
Four reflected 140-tap filters x two source emitters are expensive.
Do not publish the mode until CPU p99 on ASIO hardware, headphone
blinded same-gain A/B, front/rear clarity, timbre and DRR tests.
CI covers 44.1/48k causal delay, measured ear mirror, finite output,
legacy-default, rejection of active switching, off-silence, and
existing full source PCM/music integrity tests. Do not claim dearVR
quality merely from a synthetic regression or static source IR.

## P2 objective and CPU evidence additions
The same native offline renderer emits four 2s 48k/16bit WAVs:
FRONT/REAR x legacy light vs measured image-HRIR early reflections,
all on the identical original harmonic test input and exact direct
KEMAR. The Python evidence writer verifies direct PCM is byte-equal
before reflections, output finite/in-range, reflection difference
nonzero, and records directional-vs-light ear RMS and sample peaks.
These are **objective DSP differences**, not proof of perceived distance,
color or externalization. The informational runner benchmark compares
48k/128 3-second native wall-clock and callback p99 using the same
source and block size for light vs full measured reflections.

## P2 CPU qualification and rejected optimization (2026-11-14)

The measured HRIR P2 mode PASSED actual native C++/WAV gates on
[Action #43528783968](https://github.com/masarray/arsonmix/actions/runs/43528783968).
At 48kHz / 128 frames, the same runner 3s CPU test was:
light ER 49.992ms, directional HRIR ER 231.677ms,
ratio 4.634; directional callback p99 228.678us.
These values are for isolated Linux runner, not live ASIO.

A mirrored 2× delay-ring optimization was tested separately on
[Action #43529139184](https://github.com/masarray/arsonmix/actions/runs/43529139184):
light 49.992ms, directional 230.648ms, ratio 4.614,
p99 223.867us. The observed ~0.4% whole-window change is much
smaller than typical hosted-runner variability and costs an additional
64KiB per processor; **reject and revert** this optimization.
It does not pass an honest data-backed speed-quality tradeoff gate.

The independent test now additionally exercises moving orbit PCM
partition parity with coherent bass and reversed treble, and real
near/far wall geometry arrival-time ordering. Only testable quality
improvements are acceptable; no artificial CPU numbers or claim of
dearVR equivalence.

P2 measured per-ear reflection FIR remains intentionally offline.
A practical production renderer requires **substantially cheaper**
directional processing (such as carefully profiled SIMD/block
convolution) or a measured quality/performance policy, followed by
headphone and Windows ASIO evidence. This cannot be decided from CI
alone.

## P2 music timbre audit (candidate evidence, no perception claim)
For each of the four 48k original A/B WAVs, offline analysis
now demodulates exact 165/660/2200/4700Hz synthetic test harmonics
over a one-second steady interval, comparing light and directional
magnitude response by ear in dB. It reports RMS delta, sample peak
and a 1.5dB *engineering review flag* (NOT a scientifically validated
audibility threshold). Reflective HRTF spectral coloration is expected,
but must be assessed for musical preference and no destructive notches
with controlled blind headphones. This synthetic four-tone sample cannot
prove transparency for real multitrack music.
73 changes: 73 additions & 0 deletions docs/SONKU_ADR_P3_ZERO_SPREAD_FUSION_20261114.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
# ADR SonKu P3 — zero-source-spread linear binaural reflection fusion

2026-11-14 · draft PR #674 · OFFLINE research only.

## Research and root cause
Official dearVR PRO 2 manual §2.1 documents geometric
early reflections as critical to externalization, and §3.2 ties
reflection patterns to moving source and six wall planes.
https://files.plugin-alliance.com/products/dearvr-pro-2/dearvr-pro-2_manual.pdf
Steam Audio's implementation explicitly trades CPU against
HRTF-filtered reflection quality:
https://valvesoftware.github.io/steam-audio/doc/capi/reflections-effect.html
Neither source gives permission to copy proprietary DSP or
claims our output already matches dearVR.
The P2 renderer has 2 stereo emitter x 4 walls x 2 ears x 140 tap
FIR plus live slew per sample. 2026-11-14 CI P2 measured
~4–6x light reflection CPU on hosted Linux.

## P3 mathematically derived optimization
For exact `stereoSpreadDegrees == 0` both input emitters
travel along the identical image-source directions and acquire
identical wall path delays, gains, and 0°-elevation measured KEMAR
HRIR. Linearity gives:
`H * x_L + H * x_R = H * (x_L + x_R)`.
This holds for *arbitrary stereo*, including anti-phase material,
and does not presume L=R, add gain, or remove spectral cues.
The original stereo music-integrity side guard is upstream,
not bypassed by this optional wet-only optimization.

On each P3 room sample update both lowpass histories as before;
convolve their delayed sum only once for each of the four
image-source reflections. Advance the skipped path's HRIR filter
and delay/gain state so a live spread edit can fall back without
stale coefficients. Any nonzero spread permanently disarms
the fusion until reset. This reduces the work for zero spread,
while full 140 taps remain intact. 60° stereo spread runs
the exact P2 convolution, with no HRTF approximation.
P3 is OPT-IN through the unpublished native core accessor;
P2/alpha.23 defaults do not change; no new GUI or .smix field.

## Qualification gates
- Native 44.1/48k four input cases: true mono, independent stereo,
inverted stereo, coherent bass + inverted treble. Compare precise
floating wet outputs to unfused P2 on a rotating moving source,
assert RMS <3e-6, worst-case absolute <2e-5. Verify 60° spread
and return to 0 remain unfused after initial nonzero spread
until reset, with no stale-filter spikes.
- Same-runner 48k/128/3s P2-vs-P3 benchmark with *spread 0*,
plus a separate 60° control. No performance claims until CI.
- P0/P1 music baseline 6 WAV, P2 4 WAV and KEMAR 380
spectral checks all remain required. No release/merge pending
headphone A/B and Windows ASIO live CPU gates.

## Known limits
Optimization works only for EXACT zero source spread; it
does NOT reduce CPU on moving 60°/180° source-separated stereo.
No six-surface geometry, multi-elevation measured HRTF, headphone
individualization, head-tracking or true BRIR is claimed.

## Native same-runner performance observation
GitHub Actions #43533510068, 48kHz/128 samples/3s:
P2 spread 0 CPU **263.289ms**; P3 spread 0 **178.636ms**
(1.474x faster), p99 **250.590us -> 176.808us**.
P3 with spread 60 explicitly does not fuse (248.051ms).
These numbers are NOT full Windows ASIO/graph deadlines.

## Additional audio gate
Render four original-source 2s PCM WAVs: P2 reference and P3
fused, FRONT and REAR. Same exact 140-tap horizontal KEMAR,
room geometry, direct HRTF, 48kHz and 16-bit conversion.
Require <=2 peak PCM LSB, <2e-5 RMS, <0.05dB on
165/660/2200/4700Hz in both ears. This evaluates
sound quality invariance, not externalization superiority.
Loading
Loading