Skip to content

[SUPERSEDED — ARCHIVED] ImerSip P8 acoustic experiment (retained for A/B and IMRS migration) - #706

Closed
masarray wants to merge 131 commits into
feat/sonku-p82-sixsurface-multiband-room-20261114from
feat/grandstage-p0-sonku-p82-music-first-20261115
Closed

masarray wants to merge 131 commits into
feat/sonku-p82-sixsurface-multiband-room-20261114from
feat/grandstage-p0-sonku-p82-music-first-20261115

Conversation

@masarray

@masarray masarray commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

ARCHIVED / DO NOT MERGE. ImerSip's selected direction is PR #722, using the owner-preferred P9.1/P9.2 checkpoint #720 through PR #719. This older P8 bespoke DSP branch is retained for historical A/B comparison and IMRS v1–v5 parsing only; no source commit or branch is deleted. See docs/SPATIAL_AUDIO_AUTHORITY.md.


Audible PCM output root-cause repair — 15 Nov 2026

The owner physically tested the previous Windows candidate and reported virtually NO audible change despite scene changes. The existing design was wrong: original direct stereo fully dominated; the tiny optional roomTarget_=.14*immersion*width*(.35+.65*space) was only ~0.05 at previous defaults AND multiplied an already-attenuated complete SonKu output. Previous >1e-5 numerical tests could pass while room was functionally near-bypassed.

Real fix (not merely new GUI graphics/labels):

  • Single P8.2/P8.4 fusion PCM core now optionally exposes the same actually rendered left/right SIX-WALL EARLY and 3-band FDN LATE returns to ImerSip, without a duplicate callback, second reverb, second FIR or changes to normal SonKu PCM. ImerSip directly blends them at separately slewed audible gains.
  • Real EARLY and LATE returns are high-passed at 220/160 Hz before mixing; coherent original bass remains the only direct LF authority. Original direct stereo remains prominent. MID ambience is deliberately reduced relative to SIDE; no global gain/mute workaround.
  • Live evidence in JUCE editor: LIVE PCM DELTA -xx dB reports output vs incoming source using actual audio callback, while NO PCM CALLBACK / BYPASSED or SILENT INSERT INPUT exposes a disconnected/inaudible graph. This is a diagnostic rather than a subjective sound-quality rating.
  • Default concert preset now Immersion 82% / Venue Space 65%. Tests use the C++ initializer rather than duplicate hardcoded numbers. Existing saved .smix is not overwritten.
  • New source test with REAL P8.2+P8.4 PCM checks actual difference-to-dry on CENTER MONO AND stereo, both at default and at full concert, at 44.1/48k. Source CI run PASS at SHA 93d4589e; measured default 48k mono RMS delta −29.8 dB, stereo −20.8 dB, full concert mono −23.0 dB, stereo −17.5 dB. These are PCM difference levels, not gain or claims of audible externalization.
  • Additional REAL JUCE processBlock 48k centered music test was added on 6a3788e to reject a binary that compiles the source DSP but fails to change actual inserted FX PCM. Windows JUCE+PE+ZIP gate must complete successfully before recommending this candidate.

This remains draft because the user has rejected earlier headphone output; do not claim this will sound better without an actual, loudness-matched physical A/B against user-approved SonKu P8.4. Design/measurements.


Close Orchestra Stage — source 1.6m / early-room detail (15 November 2026)

Owner explicitly requested orchestral instruments close to the listener with more instrument detail/separation/externalization and lively room-supported, phase-coherent bass. Previous SonKu P8.4 preferred physical sound is the comparison; original P8.2 music-safe LF reference must not regress.

Committed exactly d31c44c2d7fa899099cb5b9cbe97f737221218d2 to isolated P8.2-based ImerSip branch:

  • Source is permanently 1.6 m in ALL three venues and all space settings, with existing concert room remaining large, front static width up to 180°; explicitly shown in native JUCE editor.
  • Balance now emphasizes real P8.2 six-wall EARLY reflections above the lower late FDN tail while leaving the original direct vocal MID/LEFT-RIGHT intact. Concert room amount 0.075+0.18*Space; ambience 0.025+0.10*Space. This is an audition candidate, not a measured BRIR.
  • Small (1+0.45*Space) room-linked coherent LF120 shelf gain on BOTH ears; no delayed/out-of-phase crossfeed, no separate bass engine. This is room-responsive tonal warmth, not genuine measured modal bass.
  • Added 44.1/48 kHz source regressions for near-source invariant, dry/wet scene, early > late, coherent LF monotone 45/80/120 Hz, existing original MID/SIDE, P8.4 fused direction, unmodified SonKu normal path.
  • Kept real JUCE SPSC editor/ACK, stable IMRS v1/v2 states, existing native catalog, and no changes to main GUI or SonKu P8.4.

Root cause of previous RED Windows CI: exact run 43659352042 shows all 7 Windows tests PASSED, actual EXE and packaging completed, then PowerShell ParserError: Duplicate keys physicalHeadphoneQualified during artifact provenance. It was fixed on 91fdc75 before this acoustic commit and now has a fast preflight duplicate-key guard. Do not misinterpret previous RED as a DSP or Windows compiler failure.

Current exact-commit fast tests and full Windows JUCE ZIP workflow must pass independently. See detailed field FAT. Still DRAFT—user's equal-loudness P8.4 vs ImerSip listening result needed before calling sonic improvement.


ImerSip music-first P8.4 measured fusion (owner P8.4 preferred) — 15 Nov 2026

New owner evidence: user supplied screenshot of SonKu g641b4eae and reported that P8.4 sounds noticeably better than P8.2 on real headphones. Previous PR #704 header describing P8.4 as unapproved is now superseded by this subjective preference, although calibrated FAT is still missing. ImerSip must use the P8.4 listening result as a minimum musical benchmark, not discard its progress.

Bounded code transfer, commit dd2d5f0140f185b58a4b8b04759d8baa295a2b91: copied the original 140-tap SonKuP84FusedMusicHrir.h from PR #704 WITHOUT the rejected P8.3 base or LF480 path. Only ImerSip's opt-in measured two-speaker core now uses H_center + 0.12 × HP750(H_measured_speaker − H_center) in existing FIR target updates; normal SonKu P8.2 path remains unchanged. ImerSip's processed music MID is still replaced with original direct MID outside the core, preserving vocal transparency. Source test adds original SonKu mode-OFF byte parity, directional impulse, 44.1/48 kHz LF coefficient-sum and anti-phase/vocal/SIDE/bass guard. JUCE editor remains six-control, audio ACK, v1/v2 state.

Evidence: current native PCM + actual Chromium picker CI — PASS (exact head 0cc9a4f, including P8.4 header/GUI registration and source contracts); Windows actual JUCE candidate (full compile + JUCE ACK + archive/provenance gate; artifact valid only after SUCCESS). Research design / owner reference: IMERSIP_P84_FUSED_HRIR_EXTRACTION_20261115.md.

No false quality claim: ImerSip is NOT identical to full P8.4 Music First; its uncolored dry center and controlled side-room balance differ. Do not assert better sounding/externalization or true phase-cancellation immunity across mastered music from numerical tests. Release remains DRAFT; exact Windows artifact plus side-by-side owner physical A/B (bypass dearVR/other spatial FX) required.


v2 DETAIL and editor-state reliability follow-up (source head 359c93fc)

  • Six real JUCE controls: IMMERSION, STAGE WIDTH, VENUE SPACE, BASS IMPACT, DETAIL and OUTPUT; 3 venue presets. DETAIL uses a strictly positive, identical-L/R >5.6 kHz first-order high shelf; no forced 1–4kHz voice boost, no delay-based phase-widening. Default 40%; max deliberately small (~+1.2 dB upper air).
  • IMRS v2 36-byte state; v1 32-byte restored losslessly with DETAIL OFF. Distinct state versions fail closed.
  • Native UI edit serials monotonic across editor reopen; last audio-applied ACK is latched so Save/Open cannot accidentally consume a completion and strand the editor at "APPLYING".
  • Added original-vocal MID guard with DETAIL off plus upper air/1–3kHz spectral proof, anti-phase stereo and 44.1/48 kHz buffer parity.
  • Actual Windows JUCE and physical headphone audition remain separate gates; earlier ImerSip v1 ZIP is rejected.

v2 field-acoustic redesign — voice/mono regression (15 Nov 2026)

Owner's physical audition REJECTED v1 as muffled/boxed vocal-mid, almost mono, weak externalization. Original V1 was registered but had no editor. P8.2 near-180° virtual speaker geometry from owner's better-sounding baseline is the anchor; DO NOT reuse P8.3 or inflate green source regression as listening success.

Root cause: v1 fixed P8.2 spread to 60°, spatialized all correlated music MID through centered KEMAR HRTF plus strong 6-wall room. Earlier tests only checked MID against that already processed P8.2 baseline, masking spectral coloration. That is now replaced by a source-direct reference and rejection gates.

Committed in v2:

  • Original music MID and full-band SIDE captured before P8.2 wet engine. Source MID is retained unmodified (before optional coherent bass lift), P8.2 stereo room limited to side-only contribution, bounded spatial extras; 80–180° stage (default 175°) and smaller room, fixed FRONT/no orbit. Existing SonKu P8.2 DSP itself untouched.
  • Native JUCE 840x474 real editor ImerSipEditor.h: stage map, Concert Hall/Cinema/Live Room, Immersion, Stage Width, Venue Space, Bass Impact, Output. hasEditor=true, createEditor opens through existing FX editor host, realtime SPSC/audio-applied ACK. Saved state remains isolated IMRS v1 32 bytes.
  • New objective regression against ORIGINAL dry vocal MID, not wet P8.2 MID; mono, anti-phase, stereo sides, bass 40–220 Hz, 175–180° stage, chunk parity, sample peaks. Windows JUCE CTest constructs editor and exercises state/ACK. Provenance distinguishes true native editor and unqualified physical listening.
  • Research basis: 2020 externalization review, Catic 2015 binaural reflection cues. A 175° stage knob alone does NOT prove literal external loudspeakers; measured HRTF/BRIR and listener-specific headphone FAT remain necessary.
  • v2 engineering and listening qualification

Hard blocker: Do not release as public or claim physical improvement until exact-commit Windows build, original stereo A/B and owner's headphone FAT accept it. The previous artifact #43652508958 is rejected, do not recommend.

Field fix — missing ImerSip in real FX Picker (15 Nov 2026)

The user tested the authentic Windows candidate v0.5.0-dev+g80c0af68 and the FX Picker did not show ImerSip. Root cause audited: src/NativeBuiltInFx.h guarded the registration with ARMIX_SONKU_EXPERIMENTAL_P82_ROOM, but the compiler actually receives only SONKU_P82_MULTIBAND_ROOM_RESEARCH=1. Thus full P8.2 builds and PCM tests passed while the native FX catalog omitted ImerSip. Fixed both native conditional gates to the REAL compiler define; no separate fake WebUI entry, DSP rewrite or rollback of existing mixer chrome.

Gated end-to-end: CMake -> native FX registry -> ConsoleWindow::emitCatalog -> bridge arMixCatalog -> existing source picker -> native LOAD_INSERT (no optimistic audio).

  • Windows JUCE CTest ImerSipNativeJuceP82AcceptedStateAckAndRestore now directly asserts findNativeBuiltInFx("armix.builtin.imersip.v1"), showInPicker true and constructs the actual JUCE processor through the descriptor.
  • Native PCM/compile-define GUI contract fast CI: pass run.
  • Actual compiled /preview/ native picker Chromium smoke: one ImerSip row, keyword search, click -> real LOAD_INSERT intent, no speculative FX instance; screenshot artifact — PASS.
  • PASS full Windows x64 native JUCE executable and catalog-registered ImerSip processor, full focused CTest/ACK/restore, GUI PE provenance and ZIP verifier: run #43652508958, exact head 29ac1ddee079e6dcfe43e495aeeda20b7d82ee53. Download ImerSip P8.2 experimental Windows portable artifact (expires Dec 5, 2026). Native FX picker visually tested in independent Chromium. This is a physical audition candidate, not a public release, commercial sound parity or hardware FAT certification.

ImerSip — SonKu P8.2 music enhancement with no added phase cancellation

Owner requested FX brand ImerSip (replacing P0 research name GrandStage), preserved P8.2 bass, more alive/premium musical ambience, and no hollow/quiet stereo regression. This DRAFT stacks on research SonKu P8.2 #695, NEVER on rejected P8.3 #701 and never replaces main/approved preview without FAT.

Implemented (actual native source, not just plan)

  • Real P8.2 music-safe transparent-bass / six-surface 3-band room is one native DSP core, fixed front, no rotation.
  • A coherent positive low-frequency shelf per ear (1 + b LP120, 0..+0.85 dB), exact-same L/R coefficients, no delayed bass copy; a positive high-pass SIDE shelf (1+k HP680, k<=.20) leaves MID mathematically untouched when bass is not boosted.
  • Per-sample ORIGINAL stereo SIDE detector prevents widening centered mono voices, retains negative-correlation and anti-phase stereo energy, independent of block boundaries. The peak guard suppresses only extra enhancement, never the P8.2 baseline PCM.
  • Native JUCE processor, stable armix.builtin.imersip.v1 catalog identity, 32-byte IMRS state, realtime SPSC set/completion/accepted snapshot (no editor yet), opt-in research flags only.
  • Actual P8.2 comparative PCM simulation/tests: 44.1 & 48 kHz; 3 venues × 6 mono/anti-phase/hard-panned/partial-coherence fixtures; no-added stereo MID loss; bass no-drop 40/70/110/160/220Hz; 64/128/8192 processing parity; bounded sample peaks; full native C++ fast CI.
  • Added isolated JUCE native state/ACK regression target in P8.2 research CMake gate (Windows validation pending).

Important technical honesty

Perfectly anti-phase L/R material by definition sums to zero in mono; no transparent enhancer can make it mono-safe without changing input music. The enforced invariant is no EXTRA mono phase cancellation from ImerSip, plus bass preservation vs the exact P8.2 stereo PCM. Full mastered-music proof, loudness matched listening, true peak, p95/p99 CPU and physical ASIO/headphones are still outstanding.

Current status

DO NOT MERGE/RELEASE until P8.2 stacked history is reconciled with newer main GUI, Windows JUCE test passes, approved /preview/ is used for real editor+ACK, and owner physical A/B accepts quality.

Technical math, validation design and objective vetoes: IMERSIP_P82_PHASE_SAFE_IMPLEMENTATION_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: a7a399d
Status: ✅  Deploy successful!
Preview URL: https://5dc2d261.arsonmix.pages.dev
Branch Preview URL: https://feat-grandstage-p0-sonku-p82.arsonmix.pages.dev

View logs

@masarray masarray changed the title GrandStage P0: P8.2 music-first fixed-front concert/cinema built-in DSP (research) ImerSip P8.2: phase-safe premium concert/cinema music FX with coherent bass (research) Nov 14, 2026

Copy link
Copy Markdown
Owner Author

ImerSip UI checkpoint follow-up — exact current head 2374a0b2169a8047947232826ee9787b03f68998

  • Audio acceptance remains frozen at checkpoint/imersip-180deg-approved-audio-20261115 / 30f3e333; compare_commits verified no DSP, HRIR, processor, presets or audio graph changes. Only src/Main.cpp, ImerSipEditor.h, two regression test files and the checkpoint note were changed.
  • The first full Windows run on intermediate 7c32b357 COMPILED but had a genuine red CTest: imersip_native_state_test.cpp:49 still asserted the obsolete 840×474 size. Repaired on 2374a0b2169a8047947232826ee9787b03f68998 to assert actual ImerSipEditor::editorWidth/editorHeight = 980×560 and Power exists, no test disabled.
  • Corrected the per-insert power lease to parse native channelId/epoch/revision as approved decimal-string IDs, not unsafe juce::var numeric casts; power toggle emits accepted mixer snapshot for the same native rack.
  • Native fixed-size editor canvas now cannot be arbitrarily resized into overlapped control positions; other VST3 editor resize behavior is unchanged.
  • Exact-head CI: ImerSip PCM + FX Picker SUCCESS and UI smoke SUCCESS. Windows JUCE / real popup / CTest / ZIP at current HEAD PENDING, do not distribute earlier artifact as current UI.
  • Explicit owner FAT still needed: click active rack ImerSip, X to hide, same rack click re-opens exact original instance (no duplicate), toggle POWER OFF/ON with same rack readback, verify meters/audio, rapid Save/Open preserves settings, and A/B resets on hide. Do not claim complete until this and Windows ZIP pass.

Copy link
Copy Markdown
Owner Author

OWNER APPROVED 180° SOUND CHECKPOINT — GUI reopen/power/layout repair (16 Nov 2026)

Preserve accepted audio without altering a sample: immutable checkpoint branch checkpoint/imersip-180deg-approved-audio-20261115 at previously approved SHA 30f3e333368cc3ca97b568c91f05a63f07e80980. The owner now finds ImerSip spatial rendering heading in the right direction. Do NOT change HRTF, M/S, reflections, P8.7 air/body, preset, or PCM graph. This milestone is strictly native host/editor UX.

Reported bugs: (1) active FX Rack ImerSip cannot reopen on clicking the FX slot, (2) no in-editor power/bypass like other built-in effects, (3) cramped/misaligned JUCE editor.

Fixes across six forward commits to PR #706, current exact HEAD c4e5791376af9afba9e296cb89c7035e84e373de:

  1. src/Main.cpp routed FX editor now retains a verified ChannelId/lane + slot + instanceId + epoch + revision instance lease and the same native PluginEditorWindow; titlebar X hides the window and clears transient Original A/B, while a verified same-slot click re-shows the existing editor. For replaced/reordered/retired instances, destroy GUI before lease then reacquire graph-authoritative owner. No separate DSP instance/audio callback.
  2. In-editor power binds to native bypassNativeVst3 / setDryRoutingLabFxBypass using exact epoch/revision/token; reads the insert's own atomic bypassRequested and uses the host's existing fade. Fail closed until bound; no new local DSP bypass state, and channel-wide master FX power remains independent. Native rack UI is updated from accepted snapshot, not an optimistic fake.
  3. engine/builtin_fx/ImerSipEditor.h now 980×560 aligned 4×2 controls with larger front-stage top view and clear spatial/tone labels, separated venue/Original A-B/Music Default, power top-right, readable status; no large cards. Follow-up commit changes ambiguous FX POWER state to INSERT ON / INSERT OFF, explicitly distinguishing per-slot bypass from rack master.
  4. Native JUCE CTest now checks layout bounds and correct editor size, host-power-disabled before binding, token-host binding and on/off readback, visibility false→true preserving same JUCE editor pointer and bypass state, master independent, and a new editor requiring a new host binding. It also continues accepted settings, realtime ACK, and .smix v1-v5 state checks.
  5. Unmodified approved PCM: no changes to ImerSipDsp/PhaseSafe/ConcertStage/SpatialCore; SonKu ordinary also untouched.

Research rationale: JUCE's official AudioProcessor API requires host-owned editor lifetime and explicit process bypass, and documents that editors can be deleted/replaced independently of the processor (https://docs.juce.com/master/classjuce_1_1AudioProcessor.html). Host per-insert bypass has a dedicated click-safe fade rather than a new processor flag.

Validation: Baseline pre-follow-up 2374a0b2 already has exact-head Windows + fast CI SUCCESS; current c4e5791376af9afba9e296cb89c7035e84e373de fast PCM and picker SUCCESS (#43760303415), Linux preflight SUCCESS; exact-head full Windows JUCE + ZIP #43760303439 running at this comment. Do not share older artifacts as the new build, merge/release, or claim physical rack reopen until full Windows and user FAT click→X→reopen→power OFF/ON while monitoring→save/open.

No changes to main; PR remains draft.

Copy link
Copy Markdown
Owner Author

OWNER FAT TONAL + GUI REJECTION — source-founded corrective milestone (2026-11-16)

Owner screenshots: current ImerSip at Music Default and maximum 180° sound was perceived as thin, mid-forward, short of soft bass and sparkling treble; preset at default was judged subjectively quieter, and controls appeared ineffective. Third screenshot shows existing Harmonice built-in UI with consistent Close / FX Power / preset dropdown / light silver control deck; previous ImerSip used isolated dark JUCE shell. The earlier 180° stereo engine was perceived as an improvement, so must remain recoverable at immutable checkpoint/imersip-180deg-approved-audio-20261115 (30f3e333).

Current candidate head: cca60ba404b5dfa98129efb2a8fbce5f7ec2adf5 (PR #706 remains DRAFT; no main change). Do NOT claim superior physical headphone quality yet.

Audio repair, scoped to ImerSip only:

  • Existing PhaseSafeEnhancer 120Hz and 5600Hz simple post-HRTF tone boosts were too slight at normal knob settings. Adjust LF low-shelf to ~155Hz, capped by strict <1.39x LF guard, anchored to ORIGINAL separate L/R music source (not a more powerful FDN tail). No new bass oscillator/compressor/reverb/second audio callback.
  • A first pure source-HF shelf was shown by PCM tests to decrease mid coherence: 11kHz gain only 1.0834, 1.5kHz relative 0.9566. Reject that implementation.
  • A first +4dB shelf overstated 11kHz ratio 1.517; reduce and use an RBJ biquad high-shelf of +3.0dB at ~7.2kHz designed in prepare() and identical for each ear. This preserves measured HRTF temporal relationships without a broad 1-4kHz shelf, and avoids expensive coefficient design at callback time. Existing headroom/peak guard unchanged.
  • Music Default Bass Impact now 70% (was 50), Air Detail 60% (was 40), while center, width, depth, room, stereo, session v5 and 180° KEMAR geometry remain unchanged. Saved sessions are not overwritten by defaults.

Real fixture source tests at 44.1/48k, source + Chromium PASS #43764757972:

  • 80Hz Bass 0→100% amplitude ratio 1.2817x; 11kHz Air 0→100% 1.3674/1.3628x; 1.5kHz mid ratio 1.0005x with Air change. Existing anti-phase/mono/180°/40–220Hz bass peak guards and each control endpoint still passed.
  • A mixed stereo fixture with Output held at 0dB reported default RMS ratio 1.13185/1.13591x, full controls default Output=0dB 1.25801/1.26542x relative to original. This refutes lower numerical RMS for this fixture only; NOT measured LUFS or a subjective verdict. Do not add blind auto-gain merely to inflate loudness.
  • DSP now exposes actual output energy vs input energy to editor; footer displays STereo original-side ratio and LEVEL dB instead of misleading PCM Delta as quality score.

GUI / efficient integration:

  • Native JUCE header now aligns visually with existing Harmonice motif: compact logo/subtitle, factory Preset dropdown (Custom, Music Default, Warm Hi-Fi, Grand Concert, Open & Sparkling, Studio Reference), host-owned INSERT ON/OFF, and a functional CLOSE that delegates to existing JUCE owner hide lifecycle. Silver/light control deck, metallic rotaries, mint active accents, no bulky cards.
  • Factory presets go through existing ImerSipProcessor::queueSettings SPSC + accepted ACK and same .smix v5. Manual edits mark Custom. Rejected ACK resyncs old accepted settings. Preset does not invent source, bypass authority, or second DSP.
  • Windows native JUCE test adds combo/close existence, nonoverlapping header, selecting Warm Hi-Fi and Music Default has real PCM-applied ACK/readback; hide/reopen and rack-master-independent host power coverage remains.
  • This is a factory dropdown, not yet user-named Save As; do not claim all Harmonice's WebView2 user-preset features were migrated.

Outstanding gates: exact-head Windows JUCE full CTest + portable ZIP #43764758030 in progress after Linux source and Chromium PASS. UI native screenshot and owner same-headphone level-matched A/B still required. Do NOT merge or release. Physical owner FAT must test bass warmth, vocal cleanliness, treble sparkle, full 180° stereo, preset audition, power OFF/ON and close→reopen from the rack. Existing 180° checkpoint must stay untouched.

Copy link
Copy Markdown
Owner Author

2026-11-16 — Windows RED root-cause repair (ImerSip v5)

Reconciled live PR head cca60ba404b5dfa98129efb2a8fbce5f7ec2adf5 against main; this remains an isolated draft branch, not a main release.

Verified failure: Windows run 43764758030, job 132967745417, failed while compiling the actual ARMixConsole at engine/builtin_fx/ImerSipEditor.h:448 with MSVC C2513 (const auto without variable). Root cause: factory-preset comparison named its local lambda near, a legacy Windows SDK macro which expands away; native PCM and Chromium source tests didn't compile that editor with Win32 headers.

Code fix: ffbf57e56afea5be550d11919d1f822830898045 renames the lambda to equalControlValue, preserving exact preset comparison logic, plus a fast CI macro-collision regression. f53f8f94b314c2232b9d90859e00bd991abca037 corrects guard regex escaping. Comparison against the failing head confirms only TWO files changed: ImerSipEditor.h and imersip_fx_picker_contract_test.py.

Verification: exact-head run 43771740060 fast PCM/GUI preflight PASS on f53f8f9; independent ImerSip PCM / picker / SonKu JUCE checks PASS. Actual full Windows x64 GUI compile, focused JUCE CTest, artifact provenance, and portable ZIP packaging still IN PROGRESS at this checkpoint. Do not use an older artifact or claim Windows fully green until that run ends SUCCESS.

Acoustic qualification remains separate: neither source PCM nor C++ compilation proves preferred physical headphone tonality, bass weight, sparkling treble, or premium spatial externalization. Preserve owner-approved SonKu P8.4 listening benchmark and avoid unqualified main merge/release.

Copy link
Copy Markdown
Owner Author

ImerSip 16 Nov 2026 — music source-MID fidelity / genuine early-room evidence (draft milestone)

Exact HEAD: a7a399d70ae70f68fa95ed8b1fc59d38a0c3f7e8 (stacked on existing #706). Main and SonKu P9.x isolated; no production merge.

Root cause: with STEREO KEEP=100%, the measured KEMAR DIRECT HRTF difference still injected 480 Hz+ mid-band into original headphone MID, risking radio/nasal vocal coloration. The older source test demanded >15% overall PCM change on a 370–1270 Hz fixture, effectively rewarding that coloration.

Changes:

  • Additional prepared 3-pole 3.2 kHz HP only on the direct HRTF MID difference; original direct source MID/low-frequency foundation retained. Full HRTF directional SIDE and original 2x2 measured speaker input remain.
  • Slightly stronger actual P8.2 early-reflection MID (.40 -> .62) restores physical distance cue on mono sources without increasing diffuse late MID/reverb wash. This is an audition candidate, not a measured room.
  • Split regression: original source vocal MID 500/1100/1900/2600 Hz tested room/bass/AIR OFF, independent of strong measured upper-pinna HRTF signal, full-room mono/stereo wet difference, chunk parity, phase, bass, treble, state.
  • Added docs/IMERSIP_SOURCE_MID_FIDELITY_GATE_20261116.md with physical listening hold.

Exact-head source PCM and compiled Chromium FX Picker: PASS, workflow https://github.com/masarray/arsonmix/actions/runs/43774656676.
Recorded: 44.1 kHz MID@1100 0.9887x, @1900 1.0861x; 48k MID@1100 0.98735x, @1900 1.0975x. Full room wet-null mono -22.63dB (44.1k), -21.88dB (48k), above strict -24dB veto. Direct 8.2/11.8k HRTF remains strong with controls for optional air OFF. These are objective PCM signals, not evidence of perceptual superiority.

Hold: exact-head full Windows native JUCE + ZIP provenance workflow is pending (SonKu P8.2 field run https://github.com/masarray/arsonmix/actions/runs/43774656795). Do NOT recommend older Windows ZIP as containing this change. GUI still uses separate native JUCE editor; although host-bound power/close/factory preset exist, it has NOT been visually unified with Harmonice's existing WebView2 built-in FX chrome. This must be designed/reviewed against existing app UI rather than inventing another skin. Physical equal-loudness headphone A/B vs original and owner-preferred P8.4 still required before public release.

Copy link
Copy Markdown
Owner Author

16 Nov ImerSip acoustic direction superseded by new owner-accepted SonKu P9.1/P9.2 listening baseline. Follow successor draft PR #722 for the P9 shared-native-engine integration. Do not merge P8.2/P8.4 bespoke ImerSip DSP/room shelves into P9; preserve #706 as historical A/B checkpoint and old IMRS schema reference. Production main unchanged.

@masarray masarray changed the title ImerSip P0 (DRAFT): Stereo-First MUSIC STAGE, Real Depth + KEMAR Ear Align, v5 Windows FAT [SUPERSEDED — ARCHIVED] ImerSip P8 acoustic experiment (retained for A/B and IMRS migration) Nov 16, 2026
@masarray masarray closed this Nov 16, 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