d3ro-voice/docs/evidence/2026-11-01-completion.md

9.4 KiB

2026-11-01 recovery, verification and release evidence

All dates below are KST. This report separates source tests, physical-device evidence and publication; one does not establish the others.

Recovered scope and initial state

  • Checkout: D:\workspace\D3ROVoice, sole registered worktree, branch main.
  • Initial local, Forgejo chanpaca/main, GitLab origin/main: 5a8c5cd7927a65b9289a954cff280d0d0ddf3f67 (v1.15.4). Both servers advertise main as their default. Remote refs were fetched without altering the checkout.
  • Initial dirty set: 33 tracked modifications and six new files. Meeting recording import (desktop/mobile, shared IPC/types/locales), fresh-onboarding routing, Android signer parsing, tests and maps were all retained. Original file copies, SHA-256 manifest and binary patch are under test-results/recovery-2026-11-01/initial-*.
  • No initial commits dated November 1. October 30 typing-race work and October 31 terminal/modifier/suggestion fixes were already committed through v1.15.4; they were not treated as new uncommitted work.
  • Today's earlier mobile evidence: .2 first-run gate passed; later .3 first-run auth and retry failed. The later artifacts must not inherit the earlier PASS.
  • Local session-file inventory is retained in session-inventory.json and recent-resumed-sessions.json. The root Codex session is 01a18a1c-4d07-7e31-90ca-70e542c03286; the latest local Claude log was October 28, and memory/handoff-latest.md was an August 19 historical handoff. A worker cannot query the coordinator-only Switchboard session inventory, so the local inventory is not asserted to enumerate every coordinator session.
  • Historical remote branch origin/seoro/project-overview-report, 9cd965d434acc399a661ce697f2e8ce2300c537c (April 13): one unique commit, 316 main-only commits at recovery. Its hidden titlebar/32px safe area/thin scrollbar/cloud i18n patch predates the current hiddenInset/custom Windows shell, OverlayScrollbars and translated cloud settings. Retained unmerged; blindly merging it would restore obsolete shell behavior. No additional local branch or worktree requires integration.

Implemented and recovered work

  • Desktop media picker → existing ffmpeg/STT pipeline → meeting transcript and normalized WAV, without duplicate plain history. Progress/cancel and existing minutes-generation entry points are wired. Mobile meeting creation routes to the durable meeting import pipeline.
  • Added import account-generation guards before and after WAV copy and after transcription. Before-copy, after-copy and A→B→A switches previously saved results; all now reject and remove their generated audio.
  • Retained the onboarding route-before-persistence fix and Build Tools 37 signer parser, including their tests.
  • Restored reproducible typing A/B after the fix became HEAD: default baseline f49acf01057cc100b9caeace6eada26cf7bf4b54, optional --baseline, explicit commit/hash evidence. No checkout switching is used.
  • Found public runtime feed still at 1.10.0 despite newer CPU-budget/UIA source. Sidecar minimum is now 1.16.0. The runtime-inclusive release rebuilds and smoke-tests the actual executable, publishes runtime/portable indexes before the installer, and avoids dirtying the tracked Scoop manifest after tagging. CI skips already-published runtime/portable versions.
  • Read-only release review found that rebuilding on resume could destroy bytes already partially uploaded, and checking only immutable indexes missed unfinished latest aliases. Added a pre-upload checkpoint with source tag commit and SHA-256 fingerprints. Resume reuses verified bytes through the idempotent publishers; absent/mismatched checkpoints preserve output and fail closed. CI checks both immutable indexes and latest aliases, and no longer treats network errors as an unpublished version.

Verification

Evidence root: test-results/recovery-2026-11-01/ (already ignored by the repository; no ignore changes were made).

Check Result and scope
Initial npm run verify:all PASS: desktop 2,768 tests, core 572, API client 30, mobile 87 suites / 599; workspace/mobile typecheck and lint; zero design-token violations.
Intermediate full runs Preserved five failures in verify-all-final.log: three positive runtime fixtures still advertised 1.9.0 after the minimum was raised; two voice-test import hooks timed out under concurrent load. Updated only positive runtime fixtures and bounded desktop workers to four. verify-all-bounded.log then passed desktop/core/API and lint/typecheck but retained three mobile 5 s timeouts. Bounded mobile workers to four as well, without increasing any timeout. Jest worker configuration.
Final scope checks Desktop 2,771/2,771, core 572/572, API client 30/30, workspace/mobile typecheck and lint PASS in verify-all-bounded.log. Mobile 87 suites / 599 tests PASS, 6.126 s in mobile-bounded.log. Design tokens 0 violations in design-final.log. These are composed scope passes; the intermediate full-command failures remain retained.
Account-switch RED → GREEN Three unwanted-persistence failures reproduced, then all five meeting-import use cases pass. Logs meeting-account-red.log, meeting-account-green.log.
Real MP3 / M4A integration 2/2 PASS using actual ffmpeg conversion, SQLite and WAV headers; STT response is a test double. meeting-media.log.
Typing source A/B Pre-fix 2 PASS / 15 FAIL; current 17 PASS / 0 FAIL. All three next-Hangul-word corruption cases fail before and preserve text after; unchanged-input control passes in both. test-results/typing-load/race-proof-2026-11-01T10-22-36-247Z/report.json. Three services are overlaid; dependencies/tests are current. Not native keyboard/IME/load proof.
Python sidecar suite 27/27 PASS, including CPU headroom, missing-model fail-closed behavior and cancellation. sidecar-tests.log.
Runtime feed and publication policy 44/44 PASS including five checkpoint cases: same bytes, altered bytes, missing files, source mismatch, path escape, runtime/portable parts. Old 1.10.0 feed rejected and 1.16.0 accepted. publish-final.log. Runtime install/provisioner 26/26 PASS.
Release metadata / secrets 18 negative cases PASS; source credential scan PASS. release-metadata.log, secret-scan.log.
Packaged local engine Separate-process startup/health PASS. Existing local CUDA large-v3-turbo transcribed an offline synthesized 8.498 s sentence exactly; HTTP 200, reported processing 625 ms. sidecar-smoke-before.log, local-stt-runtime.json. No cloud inference was used. This was a warm loaded model, not cold-load overlap.
Android artifact Build Tools 37 verifier and its self-test PASS. Current .3 APK: com.d3ro.voice, 1.15.4-e2e.20261101.3, code 2026110103, 147158575 bytes, SHA-256 937d2a0b3fb6b810516f07e91dd97ffe99d34c5885a72f21805a51e8a408027f; installed Fold6 identity matches. android-artifact.json.
Fold6 current runtime Existing app brought forward without clearing data. Meetings → Recording opens meeting-editor-modal with title/attendees/language/template fields; rendered screenshot and UI hierarchy retained. No meeting was saved and no inference was requested. A later list-restoration assertion failed because the expected list was no longer visible; it is not counted as a navigation PASS.
Isolated Android first-run auth Created a new empty API 34 AVD D3RORecovery20261101, explicitly targeted emulator-5580, installed exact .3 APK. Initial run failed while Android's System UI isn't responding dialog covered the app; XML and log retained. After choosing Wait, actual onboarding rendered; canonical fresh-auth flow PASS, 37 s, including persistent Login after restart. This is emulator evidence, not a new Fold6 first-run pass. Maestro also logged local heartbeat file-lock errors; those were not treated as an app diagnosis.

Open gates and preserved failures

  • Windows native GUI helper returned Computer Use native pipe is unavailable / OS error 2 after the prescribed retry and reconnect. Actual high-CPU typing, model-load overlap, native Korean IME and desktop picker-to-minutes verification remain open. Electron was not launched from an agent subshell and the user's installed app/data were not replaced.
  • Latest .3 Fold6 first-run tests and retry both fail at onboarding-screen visibility; black screenshots remain in apps/mobile-rn/.maestro-output/fold6-20261101-mtg13/. Cause is not established by those images. Earlier .2 4/4 PASS in fold6-20261101-fix1/ remains historical evidence. Personal-device data was never cleared or uninstalled in this recovery. The new isolated test AVD alone used canonical clearState flows; it contains no personal data.
  • release:mobile:config fails mobile_release_config_invalid:firebase_expected_project_id_missing, also after loading root .env. No production APK/AAB/Play release is claimed. External OAuth, Play purchase/restore, production SSV and file-to-minutes tests remain open. No paid tests or promotional messages were sent.
  • Physical-device import evidence covers the current installed local test build. The planned desktop 1.16.0 release does not imply mobile 1.16.0 was installed or published.

Release and final state

Release execution, public hashes/signature, final commit/remotes and the final clean-state check are appended after the release finishes. Until then this section is pending, not deployment proof.