[sdevil7th]

Recording and editing the supplied 31-track project exposed tiny-block processing costs, excess waveform work, source/render inconsistencies and MIDI/mixer bugs. This PR addresses those paths and packages a separately opened, offline Support application. Support stays inactive until the user opens it and requires the DAW to be closed before collection or analysis.

The owner accepted the measured 8-sample limitation as non-blocking for PR #36 on 5 November 2026, and asked to stop further optimization. The measurements remain failures; no timing threshold or test assertion was relaxed. This decision does not establish that further optimization is impossible or guarantee dropout-free recording.

Changes:

• Prepare recording storage before capture, use bounded recording FIFOs, and reduce repeated source/track work at tiny buffers while preserving deadlines and fault handling.
• Prepare native NAM input/track effects for the actual device capacity, including insertion after device startup, reducing cabinet FFT work without added cabinet latency. Preserve other processors' preparation reserve.
• Bound waveform work to the viewport; improve recording previews, tempo analysis, mixer behavior, linked crossfades, project-rate copies, joined MIDI notes and selected-clip export. Preserve the fixes shipped in 0.1.08.
• Package Support for Windows, macOS and Linux without requiring Python or terminal commands. Collect bounded existing logs, filtered audio/plugin settings and optional crash reports; inspect audio/MIDI exports offline. Users review and share archives themselves. No automatic collection, upload, service or startup task is added.
• Match incident module IDs against real PDB GUID/age, Mach-O UUID or ELF build IDs. Keep current-install and incident identities separate. Archive matching native symbols and frontend source maps at build time, outside the runtime package.
• Compile optional callback timing, ARA logging and diagnostic state out of ordinary builds. Remove routine heartbeat/fault polling; the Windows crash helper waits on events. Timed hang monitoring is restricted to explicit diagnostic tests and final Quit. Keep crash capture and functional audio/MIDI/ARA behavior.
• Keep exactly neutral EQ coefficients free of recursive fused-math residue; add a strict exact-PCM regression without relaxing the spectral export limit.
• Use bundled Vorbis for offline OGG exports so system FFmpeg builds without libvorbis work, retaining bounded cancellation, finalisation checks and atomic publication.
• Qualify installers/MSIX, Linux packages and standalone Support through nonpublishing workflows. Add an actual Linux executable/symbol/source-map archive check.

Validation at 074cc2e7ff65b72e8eb90c0e9cfaba23cf18f53c:

• Verify run: all 25 jobs pass. The Mac browser rerun passed 295 tests with 33 skips after the first attempt encountered a Chromium promise-collection error during test setup. The same commit was tested; the original failure trace is retained. No application assertions or timing thresholds were relaxed.
• Windows: pass. Native and installed-package checks cover protected normal/safe startup, detached windows, engine/export/FX tests, installer installation/uninstallation and bundled Support. The C++ build has zero compiler warnings.
• Windows App Certification Kit: pass. Certification job and report: overall PASS, exit 0, all 13 required tests pass, no required findings, published: false. Application count (1 finding) and blocked executables (179 findings) are optional failures retained in the report. MSIX SHA-256: 00b1c8e86a35ef809e007812a73b694b0266de9444df00ad67c221c1c55bffc6.
• macOS native: pass on Apple Silicon and Intel, with zero C++ compiler warnings. Native OGG quality 0-10, mono/stereo, sample-rate conversion, Unicode paths, cancellation and existing-output protection pass. The Apple Silicon export report has 429 passes and one explicitly unsupported Windows-only plugin-isolation check (not_asserted). Neutral-EQ PCM error is exactly zero; spectral route error is 1.862645149230957e-9, below the unchanged limit.
• Finished universal DMG: pass on both architectures. Apple Silicon installed report and Intel installed report identify the same DMG SHA-256, d0bb9bd0831c7336c41f0d000ee298b7f9b3ed309aeed5e661fb78ca5e7b2a9a. Each passes all 88 window checks, frozen Support, Finder launch and ad-hoc signature integrity. Notarization/Gatekeeper are explicitly not_asserted.
• Linux: pass. Native/package job: warning-free C++ build; actual executable/symbol/source-map archive matching; installed DEB normal/safe startup, detached windows, render/export and engine checks; final AppImage launch/lifecycle and ABI checks. Installed binaries retain build IDs while debug sections stay in the separate archive.
• Support and tooling: pass. Standalone Windows/Linux/Mac Support packages and universal-Mac assembly/Intel execution, Python tooling, and all three platform AI-runtime checks pass. Support remains a separately opened offline application; collection and analysis require the DAW closed.
• Local follow-up: pass. Warning-free Windows Release build, all 427 export checks, and six focused EQ suites. The exact-unity coefficient correction adds a strict exact-PCM regression rather than widening the spectral tolerance. Local precommit binaries retain their dirty-tree identity; they are not relabelled as clean commit builds.

Accepted hardware limitation — fail, non-blocking:

With Dota 2 running, Audient ASIO at 48 kHz / 8 samples and four NAM tracks at 4x oversampling, the tested executable recorded 3,658 late callbacks / 180,000 (mean 82.6 us, p99 190.8 us, maximum 408.7 us). Worker misses cover a different interval and are not additive. Raw synthetic capture passes; engine/worker capacity fails. The separate 31-track UI result remains 19 engine misses / 21 host overruns. Scheduling attribution is diagnostic_only; physical driver xruns, monitor continuity, round-trip latency and subjective quality remain not_asserted. Exact binaries, counts and trace limitations are retained in the qualification record.

Release limits:

• Developer ID/notarization credentials remain unavailable; notarization and Gatekeeper acceptance are not asserted.
• Store account inspection remains restricted by its tag-only environment policy. WACK is not Store approval; submission, update delivery and S-mode are not asserted.
• Physical interfaces, mixed-DPI transitions, Linux FUSE/RPM installation and listening checks remain manual. Scheduling attribution is diagnostic_only; subjective audio quality remains not_asserted. No formal PR review approval is recorded.

Release notes describe the unpublished candidate. The PR is open, non-draft, conflict-free and fully green, ready to merge with the owner-accepted 8-sample limitation above. No release tag, Store submission, publication or merge is included. Private recordings, project copies and machine traces are excluded.