Skip to content

P0: Cross-browser pointer stutter on standard Windows GPU (WebView2 + web mixer) #56

Open

Description

@masarray

Milestone B verified triage — 2026-11-17: P0 / gui / defect / needs-repro. Classification ≠ completion of product feature or physical FAT.

Scope / release boundary: Old GPU stutter evidence; severity conditional on current alpha.26 reproduction.

Exit criteria:

  • Capture repeatable GPU/WebView2/Firefox frame time p95/p99 and pointer jitter on exact alpha.26 and stated Windows GPU/driver at idle and active meters.
  • Fix only measured compositor/input bottleneck with browser/narrow screenshot and no native DSP or excessive repaint regressions.

Canonical issue ledger. Preserve historical evidence below; no speculative merging or rewriting approved audio.


Current engineering checkpoint (2026-10-26)

Symptom

Pointer movement outside ArSonMix is smooth, but mere hovering/moving over mixer surface visibly stutters even without dragging. Previous #54 compositor meter + drag event optimization did not resolve the symptom.

Historical baseline evidence (main at 8d5f974; newer fixes listed above)

  • web/app.js: pointermove handler exists ONLY between control pointerdown and up/cancel. Idle fader-drag math is not executing.
  • web/app.js: 14 browser demo meters continue via requestAnimationFrame at ~25 Hz. web/styles-src/40-rail-pan.css uses permanent will-change:transform on each .meter-fill (potential memory/layer overhead).
  • web/styles-src/10-console-legacy.css: stacked hover filters on fader caps, PAN thumb, inserts/FX power and strip labels; composite brightness()/contrast()/drop-shadow() chains and !important cascade on painted PNG.
  • site/preview/index.html embeds mixer.html in an iframe. Test the direct mixer page https://masarray.github.io/arsonmix/preview/mixer.html to isolate embedding.
  • User currently reports Windows desktop Firefox. CSS/WebRender or GPU driver impact unknown without profiling. Native ASIO/WASAPI/DSP are outside this work.

A/B profiling before production changes

Repeat identical pointer sweeps over the FX inserts, fader caps, PAN thumbs and a control region OUTSIDE mixer. Collect frame interval median/p95/max, fraction >33ms, paint/recalculate-style, layer/memory and long tasks:

  1. Baseline (all effects and meters).
  2. Pause meter JS updates without removing mixer layout.
  3. Temporarily disable hover filter/drop-shadow transitions on channel controls but preserve Button 19 ON/OFF styling.
  4. Remove permanent meter will-change hints; inspect layers.
  5. Combine 2-4, compare direct mixer.html with preview iframe.
  6. Compare Firefox vs Edge/Chromium on same Windows hardware, plus temporary hardware-acceleration A/B after noting about:support Graphics / Compositing.
    Record 10-second Firefox profiler trace using https://profiler.firefox.com/ in Firefox Platform mode; Chrome Performance+Paint Flashing+Layer borders for differential.

Decision tree

  • A/B hover culprit: remove per-hover blurred filters, use image/precomputed shadows or simple bounded overlay without changing fader graphics/behavior.
  • Meters culprit: dirty/visibility-aware repaint, measure single lightweight canvas meter layer vs 14 promoted HTML elements; native bridge stays intact.
  • Layer explosion: remove persistent will-change; never indiscriminately add translateZ(0).
  • iframe culprit: streamline wrapper or prefer direct preview.
  • Firefox-specific: inspect GPU driver/WebRender/software compositing and provide cross-browser CSS fallback.
  • No differential: inspect OS pointer polling, extensions, graphics driver, display refresh and main thread scheduling.

Acceptance criteria

  • Quantified before/after p95 frame time on affected Windows Firefox, not just headless Chromium green.
  • No changes to audio RT callback. Preserve Button 19, 13 strips, original 36×77 silver/gold PNGs, no-jump fader drag, PAN detent, CLIP/peak hold, native meter bridge and standalone/CSS parity.
  • Cross-browser regression and Windows native CI green before merge.

Vendor references

Activity

  1. masarray commented on Oct 26, 2026

    @masarray
    OwnerAuthor

    Phase 1 implemented and merged: PR #58 → main commit 1214047563402752ed2ec81be0dd4d14a86de686. Source hover filters simplified (fader/PAN/FX/name), permanent meter will-change removed, opt-in 6-way A/B lab added at https://masarray.github.io/arsonmix/preview/perf.html (available after GitHub Pages main deployment). Chromium smoke + Windows VST3 RT CI passed on the exact PR head, Codex findings fixed (foreground >1000ms stalls retained; centered vs non-centered PAN baseline restored). Standalone/WebView2 style/script parity remains green. Keep this issue OPEN: no Firefox GPU hardware trace or owner A/B report has been collected, so original mouse-jank root cause and user-hardware p95 uplift are unverified. Next gate: run baseline/optimized/hover-off/meters-off/layers-off/all-off, 2–3 10s sweeps each; report p95, >33ms, mouse move counts, and Firefox Profiler if needed. Do not claim production performance resolution solely from CI.

  2. masarray commented on Oct 26, 2026

    @masarray
    OwnerAuthor

    Phase 2 merged: PR #61 → main commit 34234bc085f74c67e3ef814974fd731e574865e3. Synthetic browser meter loop no longer spins rAF at 60–144Hz merely to reject callbacks; one 40ms timeout + one render-aligned rAF per tick, 60ms fallback on sustained late wakeups, restores 40ms on stable frames, and fully suspends/resumes on visibility. Native JUCE audio meter path unchanged. Standalone JS parity + full Chromium UI smoke and Windows native RT test CI both PASS on exact PR head. This confirms reduced scheduled callback activity by design, not Firefox hardware FPS improvement. Issue stays open until affected machine A/B lab and Firefox Profiler results determine remaining causes.

  3. masarray commented on Oct 26, 2026

    @masarray
    OwnerAuthor

    2026-10-26 — PR #63 merged: fixed LEDs + targeted zero-blur UI

    Merge commit: 8173119
    PR: #63

    • Fixed full-height green/yellow/amber/red palette, no scaleY; shared preview/native painter lights complete 7px rows using cached height, stable-root resize observer and --led-lit.
    • Eliminated transient fader/PAN PNG filters, toolbar/left-rail hover glows, selected-strip blur seams, native-modal backdrop blur and native scan-bar halo. Button 19 tactile state, Button 180 ownership, PAN center, peak/CLIP, audio telemetry and relative drag remain intact.
    • Deterministic CSS source/standalone parity; exact PR head 1eeceda36ebc9576ae5fd0c780161768cf797b72 passed Chromium UI smoke and Windows MSVC/20 CTest on 2026-10-26. Newest main engine P0-09 remained disjoint and passed its own CI.
    • Do not close this performance issue yet: Firefox hardware pointer frame pacing/painting, direct-vs-iframe performance, Windows DPI 125/150% and native device FAT remain unmeasured. No FPS improvement claim. Compare 10-second Firefox profiler traces with mixperf=optimized/diagnostic A/B modes, including hover and meter variants.

    Post-merge main CI and GitHub Pages build.json SHA should be verified separately before marking website publication complete.

  4. masarray commented on Oct 26, 2026

    @masarray
    OwnerAuthor

    P1 follow-up (2026-10-26): measure Firefox before changing renderer

    PR #63 is merged and green on Chromium/Windows. PR #68 integrates the final native-size 36×77 silver/gold faders and must pass exact-head Windows. PR #61 already schedules demo meter with a 40/60ms timer and one render-aligned rAF; idle fader-drag pointer logic does not run.

    Do not bulk-merge stale PR #60 Canvas renderer or claim Firefox performance resolution based on headless CI.

    Field A/B contract: compare /preview/perf.html optimized, baseline, hover-off, meters-off, layers-off, all-off; repeat 2–3 ten-second pointer sweeps across FX, fader, PAN and outside controls, with direct mixer.html vs iframe and Chromium/Edge on the same PC. Record Firefox version, about:support Compositing/GPU, DPI/refresh, median/p95 frame interval and >33.3ms count, profiler paint/compositor stack markers, plus exact tested git SHA. Firefox Profiler Firefox Platform recording: https://firefox-source-docs.mozilla.org/performance/reporting_a_performance_problem.html

    Diagnostic correctness: check PerformanceObserver.supportedEntryTypes before treating longtask as supported; unsupported does NOT mean zero. Probe rAF intervals are frame scheduling proxies, not paint/input latency. Never add background production instrumentation, new canvas renderer, speculative will-change, CSS blur or audio callback work without a measured bottleneck. Retain approved Button 19, 13 strips, fixed whole-row LED, silver/gold original faders and offline/WebView2 parity.

  5. added 2 commits that reference this issue on Oct 26, 2026
  6. masarray commented on Oct 26, 2026

    @masarray
    OwnerAuthor

    2026-10-26 — Firefox diagnostic milestone merged

    PR #71: #71
    Main merge: 3e8f860

    • Explicit mixperf-only recorder now returns p99/max and % frames >33.3ms in addition to median/p95, without instrumenting normal GUI. Longtask not supported by Firefox/API is null (not zero), checked with PerformanceObserver.supportedEntryTypes.
    • New headless Gecko functional smoke verified 13 native-size silver/gold faders, 14 clip/meter targets, no normal-mode profiler, meter-off modes, keyboard fader controls, and opt-in frame-sampling contract. CSS + JS standalone byte parity and Chromium UI PASS.
    • Exact PR head 6c2509a: Chromium/Firefox and Windows MSVC Native CI PASS. PR fix(gui): original unscaled 36x77 silver channel and gold master faders #68 native-size PNG merge also complete (a5313f8).
    • Issue remains OPEN: no affected-machine Firefox/WebRender compositor/paint trace, Windows DPI/refresh comparison or measured p95 improvement; do not claim hardware problem resolved based on CI. Capture the six A/B modes in /preview/perf.html and Firefox Profiler before changing normal renderer. Postmerge main Pages SHA should be checked separately.
  7. masarray commented on Oct 27, 2026

    @masarray
    OwnerAuthor

    2026-10-27 — additional real-machine report: Windows 11 Pro server-class PC, GPU disparity

    Owner reports BOTH embedded Windows EXE (WebView2/JUCE) and direct web preview feel markedly sluggish on mere pointer traversal over mixer (controls do not freeze); same GUI feels smooth on RTX 4050 laptop and RTX 3060 Ti desktop. Server PC runs Windows 11 Pro without HMI/other notable foreground apps. Concern is mainstream integrated-GPU/office-PC usability. No GPU model/driver, browser renderer, monitor refresh/DPI, or profiler capture yet; do not conclude integrated GPUs categorically fail.

    New source audit (main 2e249ca): (a) web/styles-src/20-button-180.css BTN-180 :hover:not(:disabled) applies filter:brightness(1.32)!important for many mixer buttons; 41-mixer-performance.css normal-mode filter suppression covers rack/fader/PAN and toolbar/left rail but NOT all mixer BTN-180 utility/channel pads. (b) web/bridge.js has a capture-phase global pointermove listener inside markNativeUiEdit even at idle; web/app.js drag move handler is attached only during active drag. (c) native src/Main.cpp emits arMixStatus on 12 Hz control timer. (d) browser-only demo meters use bounded 40/60ms timer + rAF, native meters update from telemetry. Rendering/hover and software GPU acceleration remain plausible bottlenecks, not verified; avoid claiming a fix from static inspection.

    Field differential: Windows 11 PC direct mixer ?mixperf=optimized vs ?mixperf=all-off (hover + demo meter + layer suppression where defined), 2–3 x 10s each same resolution and pointer route, Firefox/Edge if available. Check Edge edge://gpu Graphics Feature Status / Compositing, Device Manager display adapter vs Microsoft Basic Display Adapter, driver version/date, direct console vs RDP, task manager GPU Engine/CPU and display DPI/refresh. Collect Perf Lab mixer median/p95/p99/over33Pct and Firefox/Edge profiler (rAF only proxy).

    Acceptance: mainstream Windows 11 integrated GPU/office-PC baseline at 1080p/60Hz without pointer-hover sluggishness; objective p95 frame intervals and reduced slow-frame proportion on affected hardware, stable BTN-180 physical ON/press colors and controls, 13 strips, fixed meter, original fader PNG, native telemetry and audio realtime untouched. Candidate measured patches: eliminate passive hover brightness filter on BTN-180 while retaining state/pressed shadows; remove redundant idle native pointermove listener; change meter drawing only if A/B implicates it. Do NOT merge stale Canvas PR #60 as an unmeasured fix.

  8. changed the title [-]P0: Profile mixer-only pointer stutter in Firefox (hover filters / meter layers / iframe A-B)[/-] [+]P0: Cross-browser pointer stutter on standard Windows GPU (WebView2 + web mixer)[/+] on Oct 27, 2026
  9. masarray commented on Oct 27, 2026

    @masarray
    OwnerAuthor

    2026-10-27 — Senior GUI performance architecture audit and staged remediation plan

    Base: main 2e249cabb19b6f2932028521533ed6413dc78834. Field evidence: Windows 11 Pro server PC feels sluggish with idle pointer traversal on BOTH EXE WebView2 and web preview, while RTX 4050/3060 Ti PCs feel smooth. No hardware model, compositor trace, driver info or before/after benchmark. Treat the shared GUI renderer as a major suspect, not a confirmed sole cause.

    Code-backed findings and confidence

    1. P0 / high-confidence real work on hover: 20-button-180.css defines seven-layer optical shadow including inset 0 1px 20px for buttons, filter:brightness(1.32)!important on BTN-180 :hover, and transitions for filter (120ms), shadow (90ms), color (150ms), transform (75ms). There are many duplicated channel utility/MUTE/SOLO/rail/toolbar keys. This can create costly dynamic paints or composited layers on basic GPUs; actual computed style of each class on affected browser must be inspected since 41-mixer-performance.css overrides some toolbar/rail/FX hover, but not the BTN-180 utility/channel buttons.
    2. Legacy cascade: 10-console-legacy.css still contains 26 selector blocks with :hover and 61 blocks with box-shadow including historical PNG filters/drop-shadows/LED glow. Many are deliberately overridden by later source-owner modules; do not delete wholesale without getComputedStyle and reachability audits. In normal mode 41-mixer-performance.css correctly removes most fader, PAN, insert and selection blur. Static one-time gradients/bitmap materials are not the same as an active per-hover filter cost.
    3. Native event path: web/bridge.js installs a document-level capture pointermove in markNativeUiEdit EVEN without dragging (along with down/up/cancel/wheel/key). It only updates controlEditUntil for CH1/master/pan or active drag; separate web/app.js only adds its drag pointermove between pointerdown and pointerup. Safe P0 candidate: eliminate idle bridge pointermove and rely on down/up + in-flight drag in syncNativeMixerControls, with explicit long-drag/late-ACK regression. No changes in audio callback.
    4. VU paint: browser demo uses bounded 40/60ms timer+rAF (not full refresh-rate animation); native arMixStatus arrives from existing ~12Hz JUCE control timer. setLedMeterLevel caches heights and quantized 7px LED rows, but writes --level on every raw level change even when illuminated row is unchanged. Need computed-style dependency test before suppressing fallback --level, and only write changed row/display state. Never slow or falsify native meter semantics.
    5. Measurement defect: mixperf=meters-off, layers-off and baseline share CSS selectors enabling historical hover/drop-shadow and will-change:transform; all-off does not disable BTN-180 hover brightness and in WebView2 browser-demo meter switch has no effect on native telemetry. Hence existing preset A/B does NOT orthogonally isolate the expensive features. rAF interval sample measures scheduling, not Windows cursor latency, paint cost or end-to-end input response. Fix diagnostics BEFORE declaring a source confirmed.
    6. Hardware gate: Check renderer engine/GPU acceleration/driver and physical-monitor-vs-RDP; keep hardware acceleration enabled normally. Native WebView2 uses same Edge engine, so common EXE+browser lag narrows to shared GUI/OS graphics. Do not infer integrated GPUs are inherently insufficient.

    Architecture, precedence, no trial-and-error

    Phase P0-A: Orthogonal diagnostic gate (no production visual changes). Extend opt-in mixperf with independent test dimensions: BTN-180 hover brightness on/off, static 20px-inset shadow simplified/unchanged for A/B only, demo VU cadence on/off, native telemetry render on/off in an explicitly labeled diagnostic mode without ever stopping or muting the audio engine, FX hover on/off, iframe/direct. Keep modes additive AND compare to true optimized baseline; no hidden reactivation of legacy CSS. Record render version, display/DPI/Hz, CPU/GPU engine, edge://gpu, p50/p95/p99 rAF intervals, >33.3ms %, JS event counts and Paint Flashing/Performance traces. Start with server PC 1080p/60, no plugins, then 13 strips with VU.
    Phase P0-B: Safe first production cleanup. Remove hover filter brightness and its 120ms transition from BTN-180 pads while preserving exact approved resting ON/OFF 7-layer shadow, radius, palette and 1px click depression; replace hover cue with cursor/keyboard-only crisp focus outline, or no hover animation. Suppress stale but actually-active hover effects on interactive mixer controls only, after CSS computed-style audit. Remove document-level idle native bridge pointermove listener and prove no stale control ACK overwrites during 10+ second drag/wheel/keyboard. These changes need source CSS 20-button-180.css, generated web/styles.css, offline HTML parity; no global will-change, transforms/opacity hacks or audio DSP edits.
    Phase P1: Incremental meter/DOM work ONLY if profiling implicates it. Dirty-check quantized meter segments/peaks (including native stereo L/R), batch/reuse DOM lookups and avoid unchanged textContent/style writes on 12Hz native status. Cache stable references across renders safely, invalidate on renderMixer(). Do not indiscriminately replace meter architecture with old stale PR #60 Canvas; compare candidate on affected GPU with objective A/B.
    Phase P2: If compositor remains heavy. Investigate rasterized shared BTN-180 static skin asset/atlas (pixel-equivalent, no glow blur) vs live CSS, isolate software renderer/driver/remote session; consider progressive enhancement only based on measured GPU capability and never hide functionality or change public/private feature parity.

    Gates

    • All normal-state BTN-180 colors, shape, seven resting shadows, pressed 1px offset, focus-visible, ON/OFF latches; fader 36x77, 13 strips, PAN, independent LEDs, FX/Orbit context menu all unchanged.
    • Chromium+Firefox UI/standalone+CSS parity and Windows MSVC/CTest/packaging on exact head before merge. No claims about real Windows latency solely from headless CI.
    • Provisional target on representative 1080p/60 standard integrated GPU: rAF p95 ≤ 20ms, p99 ≤ 33.3ms, >33.3ms frames < 1%, no visibly slowed pointer over 10-second continuous cross-channel sweeps; verify true paint and interaction traces and correct for display refresh. These are acceptance goals, NOT measured current metrics.
    • Each PR one independently testable cause; rollback by reverted patch if p95 worsens, not by replacing the entire mixer. No partial changes to audio RT processing or VST bundles.
  10. masarray commented on Oct 27, 2026

    @masarray
    OwnerAuthor

    2026-10-27 — Render architecture audit & integrated-GPU performance plan (strategy only, no unmeasured code change)

    Affected environment: Windows 11 Pro server/office class PC, mouse pointer traversal sluggish both in web preview and native EXE, controls remain operable. RTX 3060 Ti desktop and RTX 4050 laptop smooth. Dedicated GPU must NOT be required for ordinary 1080p/60Hz mixer usage. CPU/GPU/driver/remote-session/monitor DPI are not yet measured.

    Grounded source baseline main=2e249cabb19b6f2932028521533ed6413dc78834. All CSS is built deterministically from 8 modules. Static source audit: 10-console-legacy.css ~74 KB with hundreds of retained selectors, many historical hover filter/drop-shadow and gradients; some are overridden in live normal mode and cannot be treated as active without computed style. 00-font-and-assets.css ~470 KB largely font data (startup cost, not per-frame proof). 20-button-180.css explicitly has :hover:not(:disabled){filter:brightness(1.32)!important} and seven-layer inset/outset resting shadow; transitions animate background-color, border-color, color, box-shadow, filter, transform on the shared skin. Some hover filter suppression in 41-mixer-performance.css covers toolbar/rail/fader/PAN but not all channel and utility BTN-180 surfaces. 41-mixer-performance.css also retains experimental baseline-only blur/drop-shadow for A/B; these are not confirmed active in normal mode. 10-console-legacy.css .modal blur(5px) is currently overridden by later #app .modal{backdrop-filter:none!important} in active shell; do not conflate a dormant rule with an active blur.

    Native web/bridge.js installs a capture-phase global pointermove listener along with down/up/wheel/keydown solely to mark a short UI-edit guard; web/app.js drag math itself binds pointermove only after drag begins. Native src/Main.cpp emits status every 12Hz, and web/bridge.js/web/monitor.js update status text/labels/meter DOM; repaint cost unknown. The meter uses 7px whole-cell clip-path against a static gradient, values cached; native 12Hz telemetry. web/styles-src/40-rail-pan.css intentionally uses will-change:auto, not permanent layer promotion in optimized mode.

    Sequenced implementation contract

    P0-A baseline (diagnostic-only before visual code): test direct mixer vs embedded native at same 1920×1080 display; 100% and 125% DPI where realistic; normal/mixperf=all-off/isolated BTN-180 hover disabled/LED frozen/status writes frozen. Control for RDP/direct monitor, Microsoft Basic Display Adapter, Edge GPU status (edge://gpu), CPU vs GPU tasks, GPU process compositor, hardware acceleration and driver date. Profile input responsiveness with Chromium Performance trace and Rendering paint flashing/layer borders/slow frames, not rAF-only. Three 10-second pointer sweeps per case: FX/buttons/PAN/fader/empty space; capture before changes.

    P0-B safest visual win: replace expensive BTN-180 hover filter:brightness(1.32) with a zero-filter hover treatment (or no hover modulation) while preserving EXACT seven rest shadows, OFF #b3b3b3, ON cyan/green/red/amber/blue/magenta, 3px corners and transient 1px :active travel. Stop animating filter and box-shadow globally; keep a short bounded semantic background-color transition only if measured safe. Never replace with blur, drop-shadow, will-change or pseudo-compositor hacks. Regression: computed styles at rest/hover/pressed and keyboard focus across all pad groups, 13 strips, 36×77 PNG caps, PAN and modal.

    P1-A JS pointer: remove idle native-bridge pointermove capture listener without affecting active drag: pointerdown/up/cancel and wheel/keyboard update 350ms guard, drag check protects in-flight change. Verify fader/PAN no-jump, a status update during drag does not snap the thumb, all native PR tests green. Do not add new polling or per-event layout queries.

    P1-B UI status diff/conditional rendering: keep immutable 12Hz backend status and real audio callback unchanged. Cache rendered status fields; write text/class/style only on value change, avoid full patchMixer for meter-only updates, restrict hidden Signal Check overlay to status-only updates. For meters only update quantized lit LED row and true changed peak/clip. Coalesce bursts into one display rAF where measured beneficial. Preserve real 12Hz measurement, CLIP latch and 7px fixed-color cells.

    P1-C CSS ownership/dead cascade: prove active selector reachability against actual DOM+computed style, remove superseded hover/filter/gradient legacy rules carefully (not wholesale), trim redundant transition/paint areas. Preserve default BTN-180 aesthetic and native/offline parity. No content-visibility, contain:paint or canvas experiment until popovers, sticky focus overlays, master and audio components are verified.

    P2 only if measurements prove necessary: compare original 14 HTML LED bars vs a single optimized meter-renderer candidate in independent experimental branch. Never bulk-merge stale PR #60. Evaluate bytes transferred, GPU raster, peak/clip semantics, 13 strips responsive 1080p/125%/150%, WebView2 0 VST and 20 VST, clean teardown, and p95 responsiveness. Choose the simplest renderer supported by actual evidence.

    Acceptance gates: affected Windows 11 integrated GPU/office PC, 1080p/60Hz: hover should feel native-responsive; reference p95 rAF interval ≤16.7ms desirable and no sustained intervals >33.3ms; measure paint/compositor and pointer responsiveness separately, because rAF is NOT OS mouse latency. Record improvement vs baseline and no increased renderer CPU/memory or audio glitches; Chromium/Firefox, WebView2 Windows CTest + source/standalone CSS hash parity, no native DSP changes. If edge://gpu indicates software rendering or session uses RDP, treat environment defect separately from GUI improvements, never force --disable-gpu production flags.

    Sources: https://learn.microsoft.com/en-us/microsoft-edge/webview2/concepts/performance ; https://developer.chrome.com/docs/devtools/rendering/performance ; https://developer.mozilla.org/en-US/docs/Learn_web_development/Extensions/Performance/CSS

  11. 16 remaining items

  12. masarray commented on Oct 27, 2026

    @masarray
    OwnerAuthor

    Final exact-main verified artifact — meter ease-out (#106)

    All postmerge CI on main@20222c6b72122d3adc92c92d6bee310b96aa7299 PASS: Chromium/Firefox, Windows Native MSVC + CTests + package, Pages deploy. Download directly from GitHub Actions artifact: https://github.com/masarray/arsonmix/actions/runs/40593010276/artifacts/12774973211 . Artifact ID 12774973211 (ArSonMix-GUI10CH-P012-Verified-Hardware-FAT-Windows-x64): 4,121,683 bytes ZIP, local CRC validation clean, EXE 9,118,208 bytes, SHA256 10acf3c3352c085c37a3843e3803b710a182271c271b39e49529c0a3edb3df82 matches BUILD_PROVENANCE.json, which identifies exact final main SHA. Asset/package verification is not actual physical Windows playback validation. Main nativeDSP provenance still explicitly CHANNEL_1_ONLY__REMAINING_9_GUI_CHANNELS_NOT_REAL_AUDIO. Keep #56 open until owner measured standard-PC FPS/meter ballistics.

  13. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    2026-11-05 — P0 recurrent mouse stutter audit; approved /preview/ + Windows WebView2

    Owner reports mouse movement becomes visibly sluggish inside the current ArSonMix window, smooth over other windows, with no control drag required. This is a recurrent user-visible product defect; previous Oct-2026 #56 fixes were not proven on affected hardware. Reopening rather than creating a duplicate issue.

    Current main audited: 7ad33efc4c2d6b9fbc2bf139404692414316ac21 (2026-11-05). Static source audit only. Actual physical GPU/Windows pointer profiler result: NOT MEASURED. Do not claim root cause, speedup, or resolution.

    Real product code path (not retired demo)

    AGENTS.md, scripts/build_pages.py, scripts/build_approved_native_gui.py: only approved public /preview/, generated from site/preview-v2/index.html plus approved PNG button source site/preview-buttons/assets, with native-only adapter web/approved-native-bridge.js and web/approved-native-contract.js. Old web/index.html / web/app.js and its ?mixperf measurement hooks are not the active shipping UI. Ensure any profiler, tests and fix target assembled /preview/ and the same Windows JUCE/WebView2 binary.

    Source-proven high-priority suspects, NOT measured causes

    1. Unconditional telemetry work and redundant 1 Hz burst. src/Main.cpp starts a 12 Hz JUCE message-thread timer; normal tick computes engine_.status(), emits arMixStatus and arMixConsoleSnapshot each tick. At timerTicks_%12==0, it separately computes another engine_.status() and emits an additional snapshot, plus logging. Newly added periodic accepted snapshot came from PR P0 First Sound: unlock verified hardware PATCH/ARM and stream live native snapshots #410 to fix stale device/PATCH/ARM, must not be simply removed.
    2. Native snapshot JS + DOM updates. web/approved-native-contract.js::onSnapshot validates the full snapshot and calls accepted readback; web/approved-native-bridge.js::onNativeReadback updates the hardware panel each accepted tick, constructs a serialized graph signature for native admission, and calls paintNativeMeters for unchanged graph. paintNativeMeters queries strip/meter nodes per channel. Do bounded counters and identity/version comparisons, preserve ACK→later authoritative snapshot and up-to-date device data.
    3. Meter paint even when same stepped LED is already lit. site/preview-v2/index.html::displayMeterDbfs executes fill.style.setProperty('--lit',...), marker.style.opacity and --peak-extent each call, with 14 px stepped LED quantization. site/preview-v2/assets/mixer-meter-display.css has transition:clip-path 95ms linear plus gradient/repeating masks. Test dirty-cell caching, peak-state caching, clipping/paint before selecting optimizations. Never alter actual PCM/meter telemetry or visual scale.
    4. Idle pointer costs. site/preview-v2/index.html has guarded drag-only window pointer handler, but also delegated consoleBody.pointermove (approx. line 2235) evaluating .closest('.wire-tray'), querying eligible nodes, and doing getBoundingClientRect() when hovering near WIRE. Cache geometry/invalidate on resize/bank reflow; only run proximity code in the relevant WIRE area and when enabled.
    5. CSS paint/compositor risks in approved GUI. site/preview-v2/assets/mixer-chrome-material.css still has hover filter:brightness for .strip-identity-top:hover (approx. line 447), and WIRE styles / FX row hover may recolor painted surfaces. Previous P0: Cross-browser pointer stutter on standard Windows GPU (WebView2 + web mixer) #56 2026-10-29 zero-idle-hover work was aimed at legacy web/styles-src, so its acceptance tests do not automatically prove parity for the approved current GUI. Some opt-out filters already exist: avoid blanket disabling, profile specific selectors.

    One meaningful milestone: evidence-driven GUI latency recovery

    Measure first on affected Windows PC and exact same EXE SHA: at least 3 repeat 15–30s traces, fixed window size/DPI/refresh rate; cursor sweep over blank mixer, faders, labels, WIRE, FX, and outside window. Compare (A) idle native device disconnected; (B) native device connected, meter no signal; (C) real active meters; (D) a temporary test-only no-repaint meter mode; (E) test-only no-nonessential snapshot mode preserving device/route/ACK; (F) hover effects disabled only in diagnostic build; (G) same assembled /preview/ in standalone Edge and Firefox. Compare WebView2 GPU process vs app JUCE message-thread time. Record median/p95/p99 frame interval, >33ms frame ratio, long tasks, style/layout/paint/composite breakdown, input event-to-next-paint where supported, and WebView2 native event dispatch time. Repeat with hardware acceleration A/B for diagnosis, not a forced default workaround. If there is no supported timing measurement, report unknown.

    Implement only causes supported by A/B evidence: split native structural/device state on change from low-frequency/coalesced peak data, deduplicate per-tick engine_.status() and duplicate snapshot, cache channel references and quantized meter rows, skip unchanged DOM/style writes, throttle WIRE proximity to latest RAF and invalidated layout geometry, replace hover filters with equivalent static/rest colors only where measured. Preserve same approved PNG keys, JUCE one audio callback, explicit PATCH/ARM/MUTE, exact ChannelId/EdgeId, no auto-arm, native ACK + later readback, VST3, WIRE interactions and 1366x768/1920x1080 layout.

    Acceptance: no visible pointer stutter on the affected Windows hardware while idle, live meters and during drag; benchmark comparison shows a meaningful p95/p99 and missed-frame reduction without sacrificing meter responsiveness or audio continuity. Set 60 Hz target as p95 input-to-paint under ~20 ms / >33ms frames rare subject to actual display capability, not a CI-simulated PASS. Product qualification on exact PR SHA: approved /preview/ visual screenshot/pixel contract, Chromium input tests, Windows JUCE/MSVC/CTest, native snapshot & hotplug/ACK/First Sound regressions, and physical owner FAT. No stable-performance claim without the latter.

    Avoid: arbitrary will-change, translateZ(0), global 60fps redraws, full DOM render() for every meter tick, replacing the approved skin, fake UI audio, adding realtime audio callback locks/heap, or hiding an unverified routing state as a 'performance optimization'.

  14. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    2026-11-05 — Stage A instrumentation implemented in PR #414

    PR #414 delivers opt-in instrumentation in the ONE assembled /preview/ and identical WebView2 native resource: JS rAF/handled-pointer-to-next-rAF/Long Task/snapshot-processing timing, separate gated JUCE message-thread timer/status/snapshot timing, optional bounded PowerShell ARMixConsole+msedgewebview2 CPU/memory sampler, exact-GUI Chromium test and a no-new-cross-owner CSS-overlap gate. Legacy baseline: 114 existing cross-owner selector+property combinations across 18 shipped CSS modules; known debt may be deleted, new overlap rejected.

    Crucial CSS V28 history: script injected by the deterministic Pages/native assembler, not added to historical site/preview-v2/index.html source, so all CSS V28 source hash and visual restoration tests remain meaningful. No stylesheet added or visual material altered.

    At checkpoint, explicit Chromium approved-PNG profiler smoke and CSS anti-stacking check PASS; exact PR-head Windows JUCE/MSVC and full UI suite still running. This is a diagnostic milestone, not proof the owner's physical pointer stutter was fixed. Affected Windows GPU/driver input-to-photon and actual audio hardware FAT remain PENDING. Capture guide: Stage A operator procedure.

  15. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    2026-11-05 — Stage B landed on the same single-owner PR #414 (not merged)

    PR #414 current HEAD 70632d17863d39c363fff039f74e247d727aaf01 implements runtime hot-path improvements:

    • remove duplicate 1Hz engine.status + accepted snapshot when browser subscribed (keep uninterrupted 12Hz physical readback and 1Hz AUDIO_HEALTH; disconnected browser safety evidence retained);
    • cache connected native meter nodes, dirty DOM/controls and exact --lit / peak CSS variable mutations; keep 850ms peak hold and native measured PCM;
    • memoize idle WIRE getBoundingClientRect at max 120ms age with scroll/resize invalidation; preserve drag/click control semantics and stable ACK projection;
    • no change to 18 source CSS files, approved PNG button skin, or realtime PCM; strict no-new CSS stacking gate now checks 114 existing cross-owner overlaps and 252 existing same-owner repeated declarations. These preexisting duplicates remain debt to remove with visual-parity proof, not “already cleaned.”

    At checkpoint, actual native HTML meter/WIRE Chromium functional regression, Stage A profiler, CSS ownership negative tests, immutable V28 recovery PASS. Full Chromium sweep and Windows JUCE build/CTest are running; physical Windows mouse/input-to-photon, WebView2 GPU trace and sound FAT still PENDING. Documentation: Stage B. Issue #56 stays open until affected-PC evidence actually demonstrates fix.

  16. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    Stage C — production CSS consolidation (2026-11-05, head c63c029c794903807e21ab518961ac35eff00d48)

    Implemented in the existing single-owner PR #414, retaining all previous Stage A/B native hot-path changes. The canonical Pages assembler now compiles the actual 18 original CSS owner assets for both the ONE public /preview/ and JUCE WebView2, using safe exact-selector/property/priority/conditional-context winner selection. It removes source-order shadowed declarations from generated runtime assets only, with content-addressed CSS URLs. No new stylesheet, extra JavaScript on pointermove, changed PNG hardware CSS, changed WIRE CSS, second GUI, or audio-callback modification.

    Measured compile output at exact revision: 242 redundant declarations removed, 89 of these across CSS files, 0 new override stylesheets (source CSS historical debt remains unchanged and visible). No claim that semantically distinct selectors or all conditional CSS should be mechanically merged.

    Regression evidence: full Chromium UI suite PASS, including actual baseline-versus-compiled CSS computed-style and screenshot SHA parity at 1366×768 (default, channel identity hover, hardware button hover; 1,439 elements) and 1600×900 (1,743 elements). Compiler priority/at-rule/PNG/WIRE contract, byte-exact V28 source recovery and official GUI Pages/native assembly checks PASS. Windows native build/CTest is running at this checkpoint.

    Stage C engineering documentation. Hardware WebView2 pointer stutter, GPU compositor p95/p99 and physical audio FAT remain unverified; #56 stays open until the affected PC is measured.

  17. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    Stage C integration checkpoint — exact-head 4875226

    The approved /preview/ + JUCE WebView2 build now compiles the same 18 owned CSS sources into production CSS without adding override files. Initial measurement on the early compiler removed 242 redundant declarations (89 cross-file); after adding browser-fallback safety for var()/color-mix()/vendor values, the new exact build-manifest count must be used instead (never report the older figure as final). Authoring sources remain historically byte-locked; full source cleanup is NOT yet achieved, and unsafe PNG/WIRE/feature fallbacks are kept.

    Quality work includes a second differential Chromium test across 1366x768, 1680x900 and 1920x1080, idle/hover/MUTE computed CSS styles and geometry, on top of the existing screenshot parity. Its initial failure was solely differing absolute local file:// asset roots between baseline and build; the test now canonicalizes only those test-directory prefixes, not the CSS values, image identity, dimensions or colors. Fallback safety has dedicated unit tests. Windows hardware input-to-photon evidence and actual physical audio FAT remain PENDING.

    Docs: https://github.com/masarray/arsonmix/blob/perf/p0-stage-a-approved-gui-profiling-20261105/docs/P0_STAGE_C_APPROVED_CSS_CASCADE_OPTIMIZATION_20261105.md. Keep Issue #56 OPEN. GitHub PR #414 is still unmerged, and final exact-head CI status should be checked before any promotion.

  18. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    P0 Stage C + D integrated in PR #414 — 2026-11-05

    The one approved /preview/ and native WebView2 GUI now use ONE compiled, content-hashed CSSOM resource rather than loading 18 authored CSS modules (17 fewer stylesheet requests). The original 18 source owners, accepted PNG hardware buttons and WIRE artwork keep their provenance; no new override stylesheet or visual redesign was introduced. The Stage C compiler conservatively removes 215 provably redundant CSS declarations, including 76 cross-owner in the last measured build, retaining WebView2/version and variable fallbacks. Existing authored legacy source CSS still has repeated declarations and is not falsely described as clean.

    Stage B native improvements still apply: redundant 1 Hz engine status/snapshot removed, real 12 Hz authoritative meters retained, attached-node meter caches and no-op style write avoidance, 120ms WIRE hover geometry caching without changing actual drag/ACK behavior.

    Proof from CI on integration branch: source-owner anti-stacking, deterministic Pages/Windows CSS SHA and PNG provenance, Chromium one loaded CSS/request, real MUTE, exact screenshot/computed-style parity against 18 original owner sheets at multiple viewports have passed. Final exact-head native Windows build/CTest job and actual affected-PC GPU compositor/physical input latency + ASIO/WASAPI/VST3 hardware FAT still need confirmation. The PR is deliberately NOT merged and this issue remains OPEN pending real on-device evidence.

    PR: #414
    Stage D design: https://github.com/masarray/arsonmix/blob/perf/p0-stage-a-approved-gui-profiling-20261105/docs/P0_STAGE_D_ONE_APPROVED_RUNTIME_STYLESHEET_20261105.md

  19. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    P0 GUI mouse stutter — Stages A–D integrated engineering milestone (2026-11-05)

    Exact PR HEAD: ef73e0731fdc34093e65fe44e06b6db33931b128 in PR #414. CI complete SUCCESS: GitHub Actions run 41987828893; all standalone Chromium GUI/visual parity checks, JUCE Windows executable/MSVC build, native realtime CTest, Product Qualification, experimental Windows package, approved/locked FactoryFX stage, independent ZIP/PE/embedded-GUI provenance validation and artifact upload passed.

    Experimental candidate to measure the affected Windows PC: GitHub Actions artifact ArSonMix-P0-EXPERIMENTAL-Pointer-Latency-FAT-Windows-x64, artifact ID 13397149008, under the run's Artifacts section (not a stable release). This contains the source-SHA-attested Windows GUI/ZIP for physical testing; no stable/main modifications or automatic merge.

    Runtime improvements: native 12Hz snapshot/status dedup, DOM meter/hardware dirty writes, bounded WIRE idle geometry, opt-in JS/JUCE timing; approved /preview/ and exact native GUI load ONE CSS stylesheet instead of 18 (17 fewer requests). Compiler safely removes 215 redundant runtime CSS declarations, 76 cross-file, retaining old-WebView2 fallbacks and approved PNG/WIRE assets; 18 source owner modules remain auditable. All visual and computed-style parity smoke tests including 1366×768/1680×900/1920×1080 passed.

    Hardware uncertainty remains: The original pointer-stutter issue is NOT declared fixed. Actual affected-PC GPU compositor, physical cursor-to-photon p95/p99, live VST3/ASIO/WASAPI hardware FAT and comparison to baseline remain PENDING. Start in-app 15s profiler with Ctrl+Alt+Shift+F9; independently capture native process + WebView2 CPU from scripts/collect_gui_perf_windows.ps1; review Stage A protocol. Corrected sampler recognizes real AR MIX Studio or packaged EXE process, warns if host missing.

    Merge policy: Leave #56 and #414 open until physical evidence establishes benefit and safeguards remain satisfied; green CI alone does not prove cursor smoothness. No release asset promoted to stable.

  20. masarray commented on Nov 5, 2026

    @masarray
    OwnerAuthor

    2026-11-05 — Stages A–D merged and GitHub Pages /preview/ VERIFIED (do not close P0 FAT)

    • Merge: PR P0 GUI Stutter A–D: single native CSSOM, cascade pruning, pointer hot-path + profiling #414 squash-merged successfully; main SHA 1a5ffa9309523771e950ec0dc5d3e30cef269e23. Original PR candidate head ef73e0731fdc34093e65fe44e06b6db33931b128.
    • PR exact-head qualification: Actions #41987828893 SUCCESS across approved Chromium UI, Windows JUCE build, realtime CTest, independent Windows FAT ZIP verification, and Product Qualification. Experimental Windows P0 Pointer Latency FAT artifact ID 13397149008 produced (not a stable release).
    • Main push UI qualification: Actions #41989429273 SUCCESS.
    • Pages build + deploy: Actions #41989793631 SUCCESS. Public site https://masarray.github.io/arsonmix/preview/ reachable; live /preview/build.json independently returns sha=1a5ffa9309523771e950ec0dc5d3e30cef269e23, branch=main, kind=APPROVED_SINGLE_GUI, runtimeStylesheetCount=1, removedStylesheetRequests=17, removedDeclarations=215, removedCrossFileDeclarations=76, addedOverrideStylesheets=0.
    • Stage A dormant opt-in profiler; Stage B native meter/WIRE/JS hot path; Stage C conservative CSS cascade loser elimination with fallback and approved PNG/WIRE locking; Stage D exact one compiled CSSOM in Pages and JUCE. Authored CSS has known historical duplicates; these are not claimed fully deleted.

    Important: P0 Windows physical cursor stutter has NOT been proven fixed on the affected host. The approved public /preview/ has no hardware PCM; full live device/VST3/ASIO FAT is PENDING. Keep Issue #56 open until measured p95/p99 input/paint/GPU evidence and actual hardware A/B demonstrate improvement.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions