[PCcoding666]

Original captions previously refused to start when Ollama was missing. This change makes captions the default, independent path and ships a pinned llama.cpp runtime with an optional, app-managed Qwen model download.

• Captions-first onboarding; system audio does not require microphone permission. Separate speech and translation states.
• Model progress, cancel/resume/retry, integrity checks, corrupt-model replacement and native streaming inference; no production Ollama startup, probes or recovery path.
• Preserve source-only export/history, stop-time partial captions and repeated text across independent segments; translation failures retain original captions.
• Embed/sign the runtime framework and include licenses. Update current user/developer documentation.

Validation: macos/Scripts/run_tests.sh passed Debug, Release, warnings-as-errors and 555 XCTest cases. A development-signed arm64 bundle passed strict verification and launched from Applications outside the build tree.

Draft while real GUI/audio acceptance is completed. First-run captions-first UI is observed; privacy authorizations, actual app model download/inference, offline behavior and formal Developer ID/notarized distribution remain unverified. Intel hardware is untested. The official model host timed out in a CLI probe; that is not an app-download result. Local screenshots/logs are being collected for the GUI acceptance follow-up. See docs/validation/2026-11-04-issue-25.md.

Related to #25. This PR does not close the issue or publish a release.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[PCcoding666]

Benchmark + capture check for this PR (details in #28, benchmark/engine_compare_report.md):

Runtime is not the problem. Same Qwen3.5-4B weights on this M2 Pro / 16 GB: llama.cpp Metal (Q4_K_M) vs Ollama MLX (nvfp4) are within noise — p50 end-to-end 542 vs 534 ms, TTFT 146 vs 160 ms, chrF 55.8 vs 57.6, numbers/names kept 97% vs 92%. Embedding llama.cpp costs nothing in latency or quality.

The bundled model is. The exact qwen2.5-1.5b-instruct-q4_k_m.gguf this PR ships, on the same eval set: instruction compliance 90% (vs 98%), terms kept 89%, chrF 47.7. Concrete failures: 38% → 45% translated as 3.8% → 4.5%; 五十万 as fifty thousand; several English inputs echoed back untranslated; 下周三 became "end of next week". Suggest switching the bundled model to a Qwen3.5 GGUF (4B Q4_K_M is 2.74 GB; unsloth/Qwen3.5-4B-GGUF was reachable, the official Qwen/ repo had no Q4_K_M at the time).

Capture on this branch works. Built b5b50fe with dev-run.sh: headless --speech-selftest en-US and zh-CN both clean (0 errors); GUI dual-lane with system audio (zh) + built-in microphone (en) both produced correct finals at the same time, with translation off and no Ollama installed. Not yet exercised: in-app model download and real local inference.

🤖 Posted by Claude Code