Improve runtime efficiency and native motion

[ArietidsZ]

Continuous iOS speech now splits at ten seconds and capture queues have a fixed capacity. TTS keeps native ORT KV outputs for feedback, uses exact bounded top-k selection on Android, and releases prepared-reference resources on failure. Model acquisition hashes during copying/downloading while preserving authentication before archive parsing, staging and rollback.

Native Compose/SwiftUI motion now clarifies session state, new turns, press feedback and live audio levels. Reduced-motion preferences are applied on both platforms; idle timelines, per-bar animation state, an unused status view and redundant lifecycle scaffolding were removed. The branch includes an isolated 18-second Remotion showcase using the existing Auralis brand.

Validation:

• Eight final-source full-model Swift WAVs match the released baseline byte for byte; cancellation, no-EOS rejection and failure recovery pass.
• 23 real-ORT Swift tests, eight waveform tests and 48 session host checks pass. Full-source Catalyst typecheck passes with real platform headers.
• Android: 169 tests, zero failures, one explicit benchmark skip; debug build and API 35 UI-only AVD checks pass, including 200% text, navigation and system reduced-motion return.
• Contract/Python/Swift fast checks pass. Local Android lint was blocked by dependency-cache/TLS conditions; final GitHub CI runs lint against official repositories.
• The 256 MiB APFS acquisition benchmark improves from 0.158 to 0.128 seconds. Sampling microbenchmarks and equivalence evidence are recorded in the area reports; these are host measurements.

Final commit e3d9a82 passed all three jobs in both push CI and PR CI, including native Android/iOS builds, tests, lint and the showcase source check.

Evidence and ablation: quality report index.

Physical-device performance and active-meter frame timing remain unqualified. Shared model packages remain draft. Published v0.1.0-preview.1 artifacts retain their original contents.