iOS Parity (v1.7.1) — tracking issue

[Pivii]

Meta-issue tracking the catch-up with dictus-ios main @ v1.7.1. Full analysis: claudedocs/ios-parity-gap-analysis.md (gap analysis 2026-06-11, validated). LLM/Smart Mode (iOS develop branch) is explicitly out of scope — separate future effort.

Wave 1 — Quick wins & known bugs (P0)

• Wire Home 'Nouvelle dictée' button to the recording screen #18 Wire Home "Nouvelle dictée" button
• Onboarding: replace iOS-style keyboard setup card with real Android IME enable flow #19 Android-native onboarding keyboard setup (replace iOS-style card)
• Model cards: render description + accuracy/speed gauges #20 Model cards: description + accuracy/speed gauges
• Settings → About: show versionCode alongside versionName #21 versionCode in Settings → About
• Align minimum recording duration with iOS (1.0s → 0.5s) #13 Minimum recording duration guard (0.5 s)

Wave 2 — Recording UX & engine (P1)

• Full-screen model loading overlay + mic gating during model load #22 Model loading overlay + mic gating
• Engine pre-warm on IME bind + idle release; align transcription watchdog #23 Engine pre-warm + idle release + watchdog alignment
• Animated mic button states (idle glow / recording pulse / transcribing shimmer / success flash) #24 Animated mic button states — reduced-motion/battery-saver sub-slice Respect reduced motion and battery saver in synthetic waveforms #97 merged via feat(ui): respect reduced motion for synthetic waveforms #98; remaining idle/success semantics need a design decision that preserves approved Android replacement screens
• Audit Bluetooth audio routing during recording (headset mic, media playback) #25 Bluetooth audio routing audit — deterministic built-in mic policy merged; accessory matrix remains hardware-blocked
• Privacy: redact transcription text from logs + clear stale last transcription #26 Privacy: log redaction + stale transcription cleanup

Wave 3 — Autocorrect & prediction (P1, biggest chunk)

• AOSP Patricia Trie spell correction (C++ NDK/JNI) #7 AOSP patricia trie spell correction (C++ NDK/JNI)
• N-gram next-word prediction (bigrams + trigrams) #8 N-gram next-word prediction
• LanguageProfile architecture: per-language data files + autocorrect toggle #28 LanguageProfile architecture + autocorrect toggle
• Autocorrect-on-space with backspace undo and rejection learning #27 Autocorrect-on-space + undo + rejection learning

Wave 4 — Languages (P1, builds on Wave 3 for autocorrect quality)

• Spanish language support + keyboard language switcher #12 Spanish + keyboard language switcher
• German language support (layout, umlauts/ß, collapse rules, dictionaries) #29 German support

Wave 5 — Keyboard polish (P2)

• Trackpad cursor mode: long-press spacebar to move cursor #30 Trackpad cursor mode
• Key feedback polish: per-key popup preview, accelerating backspace, key sounds #31 Key feedback polish (popup preview, accelerating backspace, key sounds)
• French adaptive accent key on AZERTY (apostrophe/accent based on context) #32 French adaptive accent key

Decided out of scope

• Whisper Turbo → Evaluate Whisper Turbo support and device gating strategy on Android #11 closed (slow even on iOS; Standard recommended)
• Live Activity / Dynamic Island equivalent → no Android need; much later if ever
• Transcription history (Transcription history with swipe-up gesture from home screen #10) → not a parity gap (absent from iOS main too); stays in the feature backlog
• QWERTZ layout → deferred on iOS too (dictus-ios#151)

iOS source releases covered

v1.3.0-beta → v1.7.1 (March–May 2026): keyboard rebuild, AOSP trie + n-grams, Spanish, AirPods/A2DP fix, Live Activity toggle, onboarding overhaul, autocorrect quality (#113), Turbo gating, LanguageProfile refactor, German, model-loading UX, cold-start fixes.

Current campaign state

All non-ambiguous, currently executable parity slices are complete. #24 retains a product/design decision; #25 retains a Bluetooth/wired accessory gate. Evidence and exact residual scope are documented in their issue threads.

[Pivii]

Overnight autonomous campaign started 2026-08-25. Initial Wave 1 audit:

• Align minimum recording duration with iOS (1.0s → 0.5s) #13 genuinely missing → implemented in PR fix(recording): reject clips shorter than 500ms #36 with 500 ms boundary tests; independent review approved; CodeRabbit pending
• Wire Home 'Nouvelle dictée' button to the recording screen #18 wiring already exists; issue narrowed to regression tests + current-build UAT
• Onboarding: replace iOS-style keyboard setup card with real Android IME enable flow #19 partly implemented; remaining gap is enabled → selected IME flow and picker
• Model cards: render description + accuracy/speed gauges #20 UI already exists; remaining gap is regression tests + Pixel 4 calibration evidence
• Settings → About: show versionCode alongside versionName #21 exact version format merged via PR fix(settings): simplify version display to match iOS format #34 and issue closed

Baseline: 99 unit tests passing locally, debug APK assembles. Hourly campaign continues with the smallest non-blocked slices.

[Pivii]

Campagne autonome — jalon 2026-08-26 00:47 CEST

• ✅ Align minimum recording duration with iOS (1.0s → 0.5s) #13 fermé via fix(recording): reject clips shorter than 500ms #36 (9eee23e) : durée minimale 500 ms, synchronisation de la lecture audio finale, tests de frontières/course.
• ✅ Wire Home 'Nouvelle dictée' button to the recording screen #18 fermé via fix(recording): auto-start after late controller binding #37 (f051a36) : navigation déjà présente ; correction du binding tardif empêchant l’auto-start.
• Validation fix(recording): reject clips shorter than 500ms #36 : 113 tests, 0 échec ; assembleDebug OK ; sabotage sans join échoue 2/2.
• Validation fix(recording): auto-start after late controller binding #37 : test ciblé RED→GREEN ; 229 tests, 0 échec ; assembleDebug OK.
• Revues indépendantes fraîches : aucune anomalie bloquante ; CodeRabbit vert sur les deux têtes exactes.
• lintDebug reste bloqué sur la base develop par la dépendance Gradle Licensee déjà connue et hors de ces diffs.
• Captures : non applicables, aucun rendu visuel modifié.

Prochaine tranche autonome : auditer #19 sur develop, confirmer le flux IME activé → sélectionné, puis isoler le plus petit manque testable sans conflit.

[Pivii]

Rapport final de la campagne nocturne — 2026-08-26 10:00 CEST

Livré sur develop

• ✅ Settings → About: show versionCode alongside versionName #21 via fix(settings): simplify version display to match iOS format #34 (30c1887) : version affichée au format iOS 1.1.0 (12).
• ✅ Align minimum recording duration with iOS (1.0s → 0.5s) #13 via fix(recording): reject clips shorter than 500ms #36 (9eee23e) : seuil minimal de 500 ms / 8 000 échantillons, rejet avant initialisation STT, et synchronisation de la dernière lecture AudioRecord.
• ✅ Wire Home 'Nouvelle dictée' button to the recording screen #18 via fix(recording): auto-start after late controller binding #37 (f051a36) : le câblage Home → Recording existait déjà ; correction de la course de binding tardif pour auto-démarrer dès que le contrôleur devient disponible.
• Aucun PR ne reste ouvert. Tête actuelle de develop : f051a36.

Validation vérifiée ce matin

• fix(recording): reject clips shorter than 500ms #36 : testDebugUnitTest — 113 tests, 0 échec, 0 erreur, 0 ignoré ; sabotage sans attente du job de capture — 2 tests échouent comme attendu ; assembleDebug OK, APK 160 769 319 octets.
• fix(recording): auto-start after late controller binding #37 : test ciblé RecordingScreenControllerBindingTest RED→GREEN ; testDebugUnitTest — 229 tests, 0 échec, 0 erreur, 0 ignoré ; assembleDebug OK, APK 160 673 397 octets.
• ./gradlew lintDebug relancé sur l’arbre exact de develop (b7e2df5) : BUILD SUCCESSFUL en 45 s. Le blocage Licensee annoncé pendant les PR n’est donc plus présent. Rapports : 0 erreur, avertissements non bloquants existants.
• CodeRabbit : vert sur les têtes exactes de fix(recording): reject clips shorter than 500ms #36 et fix(recording): auto-start after late controller binding #37 ; aucune observation bloquante après correction de la lecture audio finale.
• GitHub Actions n’a pas tourné pour ces PR vers develop, car .github/workflows/ci.yml ne cible que les push/PR vers main. La validation de cette campagne est donc locale + CodeRabbit, pas un CI Actions de branche.

Captures

Aucune nouvelle capture : les trois changements n’altèrent pas le rendu, sauf le texte de version d’une ligne. Captures de référence déjà versionnées :

• Clavier
• Modèles
• Paramètres

À valider sur appareil réel

• Align minimum recording duration with iOS (1.0s → 0.5s) #13 : essais micro avec « oui », « non », « ok », « ah », et confirmation qu’un clip < 500 ms est refusé sans hallucination.
• Wire Home 'Nouvelle dictée' button to the recording screen #18 : parcours complet Home → Nouvelle dictée → micro → transcription → copie → retour Home, avec mise à jour de la dernière transcription. Le bug de binding est couvert en Robolectric, mais ce parcours matériel n’a pas été exécuté cette nuit.
• Model cards: render description + accuracy/speed gauges #20 : calibration des scores vitesse/précision sur Pixel 4 ; l’UI demandée existe déjà, mais la preuve de mesure manque.

Priorités recommandées

1. Onboarding: replace iOS-style keyboard setup card with real Android IME enable flow #19 : implémenter les trois états onboarding IME (non activé / activé non sélectionné / sélectionné), ouvrir le sélecteur Android et tester la reprise après Paramètres/process death.
2. Model cards: render description + accuracy/speed gauges #20 : ajouter les tests de contrat/catalogue et Compose sans réécrire l’UI, puis faire la calibration Pixel 4.
3. Étendre le workflow CI aux PR vers develop avant la prochaine vague, afin que lint/tests/APK soient des checks GitHub obligatoires.

[Pivii]

Correctif au rapport nocturne et build physique prêt

L'audit post-campagne a identifié deux problèmes de processus :

1. le job nocturne avait été créé avec once in 1h, donc il n'a exécuté qu'une seule reprise à 00:48 malgré les dix répétitions annoncées ;
2. le lint annoncé vert à 10 h venait d'un worktree incrémental. Un checkout propre de develop@f051a36 reproduisait la dépendance Gradle manquante entre les tâches lint et les assets Licensee.

Correction fusionnée : PR #38, commit 37147a2.

Validation exacte sur un checkout propre de develop@37147a2 :

• ./gradlew clean lintDebug testDebugUnitTest assembleDebug : succès ;
• 229 tests, 0 échec, 0 erreur, 0 ignoré ;
• APK : 160 706 165 octets ;
• émulateur API 30 : Accueil, Paramètres, version 1.0.0 (1), Home → Nouvelle dictée et auto-démarrage de l'enregistrement vérifiés ;
• limite : la transcription x86_64 ne peut pas charger libwhisper.so, qui est ARM-only. Le test micro/STT final doit être fait sur Pixel 4.

APK pour UAT physique : /home/hermesbot/exchange/Dictus-Android/dictus-android-develop-37147a2-debug.apk

SHA-256 : e1079dda3bac5bd12834aab9e317e66fd9b620974dbdb0f6268dbfd0f57cb24a

[Pivii]

Gate Pixel physique intégré — develop@1835955 (2026-08-27)

Gate rattrapé sur le checkout propre de la tête exacte 18359550de26c8d98475a113e434f3d74a16c141, qui contient #34, #36 et #37.

Gate hôte exact-head

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug : BUILD SUCCESSFUL
• JVM : 229 tests, 0 échec, 0 erreur, 0 ignoré (30 XML)
• lint : 0 erreur (avertissements baseline/non bloquants)
• APK debug : 160673397 octets, SHA-256 bf4dbabf114a8ecbac8ba0ac5d23bb0d420d0696281227226bb408d7deaca6ed

Gate Pixel

• appareil : Pixel 4 flame, Android 15 / API 35, arm64-v8a, série 99071FFAZ00EBN
• APK exact installé depuis le checkout ci-dessus ; package vérifié versionName=1.0.0, versionCode=1, ABI primaire ARM64
• Paramètres > À propos affiche réellement 1.0.0 (1)
• Home > New dictation : le service bindé réel démarre automatiquement ; logcat confirme le FGS microphone puis AudioRecord started: 16000Hz mono Float32, et l'UI affiche l'état Recording
• clip physique court : 1920 échantillons, rejeté contre minimum 8000 avec clip too short; aucune ligne ModelDownloader, WhisperProvider ou ParakeetProvider après ce clip
• clip >=500 ms : 201600 échantillons, transition réelle vers Transcribing, téléchargement Tiny, initialisation Whisper ARM64 (NEON=1, ARM_FMA=1), puis fin normale du chemin de transcription
• aucune exception/crash/ANR Dictus observée dans les fenêtres logcat collectées

Confidentialité des preuves

La dictée longue a capté de l'audio ambiant. Sa transcription n'est ni citée ni conservée : capture/XML/logcat correspondants supprimés immédiatement, LAST_TRANSCRIPTION remis à zéro via une fixture DataStore non sensible, files/dictus.log supprimé et logcat vidé. Les seules captures conservées sont non sensibles (version, état Recording, rejet sans résultat) dans /tmp/dictus-pixel-evidence-1835955/.

Hashes des preuves locales :

• settings-version.png: b69a0c3f29698e36682e01286c94f195f03fead773ac73d3d37ecddde7606a41
• recording-auto-start.png: cea7e1a5be834bdaefaac5a28cfe12dc7fadee3653981d95a9742457682524ce
• short-policy-rejected.png: 2028778859bc6dbf6f64943e131f0a0a914ed08da8d5e1abf4c7c806e050650d
• log court filtré: 70840d03fcd7f26977e4bbdf9b6626f2e9f9232045688c6a84015836eb3c7a81

Conclusion

Le gate physique différé de #34/#36/#37 est vert sur la tête intégrée actuelle. Le test a également reproduit le risque confidentialité déjà suivi par #26 (texte brut dans les logs + transcription persistante), qui devient la prochaine tranche prioritaire.

[Pivii]

Integrated physical Pixel gate update

The develop gate for #34/#36/#37 exposed a real waveform concurrency crash after ~17 seconds of microphone recording (WaveformBars.kt:64). Root cause and physical evidence were tracked in #42; the synchronized-history repair was independently reviewed, exact-head built/tested/linted, exercised on Pixel 4 Android 15 for 23 seconds without crash/ANR, and merged as #43 (e47dffc).

Verified on physical Pixel during this gate:

• Settings version 1.0.0 (1) matches package metadata
• Home → New dictation starts real bound foreground DictationService + AudioRecord
• 640-sample clip rejects before model/provider initialization (8,000-sample / 500 ms minimum)
• 42,240-sample clip transitions to Transcribing and model acquisition
• app data cleared after runs; no audio/transcript/personal data retained

Repository CI remains absent for these branches; exact-head local gates and fresh-context review were used per current policy.

[Pivii]

Physical Pixel integrated gate — develop e47dffc

Executed on Pixel 4 (flame, Android 15, ARM64), serial ending 00EBN, using a clean detached checkout with initialized whisper.cpp submodule.

Exact-head host gate

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug — passed
• JVM: 244 tests, 0 failures/errors/skips across 32 XML suites
• APK: app-debug.apk, 160,722,549 bytes, SHA-256 7c85362c0112c3070b9e6af5311cca0a6670db9cbcfbe242b2c80fd4e08c88f2

Exact APK physical evidence

• Installed successfully with version metadata 1.0.0 / code 1.
• Settings rendered 1.0.0 (1), matching the iOS-format change from fix(settings): simplify version display to match iOS format #34.
• Home → New dictation auto-started the real foreground service after binding; UI showed recording state and logcat confirmed AudioRecord started: 16000Hz mono Float32 plus DictationService started foreground with ACTION_START (fix(recording): auto-start after late controller binding #37).
• Deterministic short attempt: physical capture returned 640 samples; policy logged clip too short (640 samples, minimum=8000), returned (No result), and emitted no model/provider initialization or transcription event (fix(recording): reject clips shorter than 500ms #36).
• Long attempt: 241,920 samples entered Transcribing, reused the downloaded small-q5_1 model, loaded whisper_v8fp16_va (ARM fp16), and completed native offline inference successfully. Privacy-safe logs retained lengths/timings only.
• App remained top-resumed; focused logcat scan found no Dictus fatal exception, process death, or ANR.

Non-sensitive evidence is retained locally at /tmp/dictus-pixel-settings-e47dffc.png and /tmp/dictus-pixel-short-clip-e47dffc.log. No audio file or transcript artifact was saved.

Residual: this gate covered #34/#36/#37 on integrated develop; it was not an acoustic-quality assessment.

[Pivii]

Physical Pixel integration gate — develop e47dffc (2026-08-27)

The deferred hardware gate for #34, #36, and #37 is complete on the exact current develop head e47dffc8e06f974cfcd29edcad47226b78735d40.

Host gate (clean exact-head checkout)

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug — passed
• 244 JVM tests across 32 suites; 0 failures, 0 errors, 0 skipped
• APK: app-debug.apk, 160,722,549 bytes
• SHA-256: 7c85362c0112c3070b9e6af5311cca0a6670db9cbcfbe242b2c80fd4e08c88f2

Physical gate

• Device: Pixel 4 (flame), Android 15, ARM64, serial recorded in the private campaign configuration
• Exact APK installed with pm install -r -t: Success
• Settings UI rendered Version = 1.0.0 (1), matching the intended VERSION_NAME (VERSION_CODE) format (fix(settings): simplify version display to match iOS format #34)
• Home → New dictation promoted the real microphone foreground service and started AudioRecord at 16 kHz mono Float32 without a second mic tap; this exercises the real activity/service binding path (fix(recording): auto-start after late controller binding #37)
• Deterministic short-sample run captured 3,840 samples (240 ms at 16 kHz): rejected as clip too short, minimum 8,000; no provider initialization or transcription followed (fix(recording): reject clips shorter than 500ms #36)
• Accepted-sample run captured 11,520 samples (720 ms at 16 kHz): transitioned to Transcribing and completed ARM-native Whisper inference successfully (fix(recording): reject clips shorter than 500ms #36)
• ARM evidence: earlier same exact-head run loaded whisper_v8fp16_va; system info reported NEON/ARM_FMA; model inference completed on device
• MainActivity remained top-resumed; filtered logcat contained no FATAL EXCEPTION or ANR in

Privacy/evidence handling

• No audio was exported or retained by the campaign.
• No transcript text is included in GitHub evidence; temporary UI XML containing live result text was deleted.
• A non-sensitive Settings screenshot and redacted-safe state/log evidence are retained only on the campaign host under /tmp.

Residual risk: this closes the requested integrated hardware regression gate. It is not a broad acoustic-quality, Bluetooth-routing, or benchmark claim.

[Pivii]

Physical Pixel integration gate — PASS on develop f224ad3

Retrospective physical gate for #34, #36, and #37 was run against exact develop commit f224ad38aa581d4a245d8aef57eff245ec98a2d4 on a USB Pixel 4 (flame, Android 15, ARM64; serial suffix ...00EBN).

Exact-head host gate

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug: PASS (BUILD SUCCESSFUL, 297 tasks)
• JVM XML: 257 tests, 0 failures, 0 errors, 0 skipped across 36 suites
• APK: 160,739,409 bytes
• SHA-256: d0187a8484ae2c65605f4f06fb104457c51473e769f322c8eedba5f906a73ddb

Physical evidence

• Installed the exact APK through adb push + pm install -r -t: Success.
• Settings UI hierarchy showed Version → 1.0.0 (1). A non-sensitive screenshot was captured locally; SHA-256 8419af07ff3c7931ad5e8573321116602570845f3285d1708c1949184f53cfcd.
• Home → New dictation auto-started through the real late-bound service path. UI entered recording (0:03, Stop), and logcat showed AudioRecord started: 16000Hz mono Float32 plus foreground ACTION_START.
• Valid clip path: physical capture produced 290,560 samples; state transitioned to Transcribing, active small-q5_1 resolved, native ARM Whisper context initialized in 200 ms, and inference completed. Transcribed content was not retained or posted.
• Short-clip oracle: a 450 ms scripted gesture produced 1,920 samples; service logged clip too short (1920 samples, minimum=8000). Assertions confirmed no Transcribing transition, model lookup, Whisper context initialization, or transcription marker.
• Crash/ANR scan: 0 FATAL EXCEPTION / app ANR markers during the exercised flows.

Privacy cleanup / residual risk

• No audio was exported or retained. All UI XML containing runtime text was deleted.
• App data was cleared after the run (removing downloaded model, preferences, and any transcription state), the temporary APK/UI dump was removed from the Pixel, and the default IME was restored to AOSP.
• This gate proves real Pixel microphone/service/native wiring and the 500 ms boundary behavior. It is a smoke gate, not a microphone-quality, noise, Bluetooth-routing, interruption, or performance benchmark.

[Pivii]

Campaign milestone — engine prewarm merged (#49)

Issue #23 is closed via PR #49, merge commit fdbdeb3924869ec8109e91180734e38cf2fc985a.

Delivered:

• downloaded active model prewarms off-main when service binds / real IME connects; no bind-triggered download
• one mutex owns prewarm, transcription, model replacement, and release
• renewable 10-minute idle lease releases native memory
• Cold/Loading/Ready/Failed state is now available for Full-screen model loading overlay + mic gating during model load #22
• 120 s watchdog retained pending larger-model timing rather than generalized from Tiny

Exact integrated clean-checkout gate on develop@fdbdeb3:

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug — PASS
• 264 JVM tests across 36 suites, 0 failures/errors/skips
• APK 160,755,793 bytes; SHA-256 7cfd0339e04ce967be31c80bc16be24eafac1244e776cc3546091885f84f6d1c
• fresh-context exact-diff review: no blocking findings; CodeRabbit status green/rate-limited

Pixel 4 Android 15 / ARM64:

• actual system IME bind prewarmed Tiny in 172 ms
• first physical dictation reused the warm context (0 reloads) and completed inference in 2,202 ms
• service/IME teardown logged native provider release
• 0 app fatal exceptions / ANRs in the exercised window
• audio/transcript content not retained; app data and temporary device files cleared; Latin IME restored

Next risk/value slice: use the newly exposed readiness state to audit #22 mic gating/loading overlay before broader keyboard polish.

[Pivii]

Campaign milestone — model loading gate merged (#50)

Issue #22 is closed via PR #50, integrated commit 5e2851b.

Delivered: one shared full-surface loading/failure UI across standalone recording, onboarding, and IME; cold/loading mic intents wait for exact Ready and start once; retry never invents microphone consent; unavailable models become actionable failures; underlying touch/accessibility interaction is gated.

Exact PR gate: clean lint/JVM/APK passed; 274 tests, 0 failures/errors/skips; fresh-context re-review approved after three blockers were fixed; CodeRabbit check passed (one import suggestion verified false against the exact compiler).

Pixel 4 Android 15 / ARM64: real unavailable-model failure UI, Small Q5 retry/loading (229 ms), no microphone start from Retry, explicit post-Ready recording, and whole-system IME bind/prewarm (260 ms) all exercised; 0 crash/ANR markers; no audio/transcript retained and device state cleaned.

Next risk/value audit: #25 Bluetooth audio routing, beginning with current AudioRecord/routing behavior and deterministic route-change coverage before any hardware claim.

[Pivii]

Integrated exact-head host gate — develop@bcba9a6

Clean detached checkout with initialized whisper.cpp submodule:

• ./gradlew --no-daemon clean lintDebug testDebugUnitTest assembleDebug: PASS
• 277 JVM tests across 41 XML suites; 0 failures/errors/skips
• APK: 160,826,265 bytes; SHA-256 060b8d07cd5383d8d09e39d0a868986eeb97e86b3a787eb32beb352350e83e05
• develop@bcba9a6 and independently reviewed/Pixel-tested PR head e1027bf have the exact same Git tree (ba1ee536...), so the merged integration contains the already-recorded Pixel built-in-mic routing evidence without code drift.

Next concrete slice: #53 (accelerating held backspace), child of #31.

[Pivii]

Pixel follow-up on current develop@bcba9a6: exact clean host gate passed (277 JVM tests, lint 0 errors, assembleDebug), Settings showed 1.0.0 (1), Home → New dictation auto-started real ARM64 AudioRecord, a deterministic 640-sample clip was rejected before transcription, and a 250,880-sample clip completed native Whisper inference with no crash/ANR. App data and temporary device artifacts were cleared. The clean onboarding run also exposed a stale first-load failure immediately after model download; tracked with repro and acceptance criteria in #54.