End-to-end latency benchmark: key press → transcription → paste, fully local #34

Existing benchmarks cover parts of the pipeline: make benchmark (encode + provider + network) and make bench-local (local inference only). Neither measures what the user actually feels: key release → text pasted, including hotkey handling, capture teardown, encode, inference, clipboard, and paste.

Add a repeatable end-to-end benchmark that:

• runs fully locally (local model, no network), so results are stable and machine-comparable
• measures press-to-paste wall time, ideally as a Go benchmark (testing.B) so benchstat can compare runs; a manual/semi-manual harness is acceptable if real key events make a pure Go benchmark impractical (e.g. feed a canned WAV through the post-release path)
• can be re-run from time to time and appended as a labelled baseline (same spirit as make bench-save) to see what we lose or gain across changes

The per-segment timings already logged to diagnostics (press_to_record_ms, paste_copy_ms, ...) are a good starting point — the benchmark mostly needs to drive the pipeline and aggregate them.

🤖 Generated with Claude Code