fix(voxtype): read final text from streaming sessions

[quanru]

Summary

• Accept Voxtype 1.1 streaming sessions as an active recording state.
• Pass --wait-file when the CLI supports it, so a caller with a separate config can wait on the private transcript path.
• Recover the completed transcript when Voxtype's streaming file mode returns exit 3 after writing the file without a .done sidecar. The normal final-result path then allows Doubao Say polishing.
• Document the current partial-text limitation.

Dependencies and merge gate

• Doubao Say stack: feat(asr): integrate Voxtype local dictation #48 → feat(settings): show Voxtype runtime controls #49 → feat(voxtype): select an installed local model #53 → this PR.
• Upstream live transcript dependency: peteonrails/voxtype#728. It proposes an atomic $XDG_RUNTIME_DIR/voxtype/transcript mirror updated after every Partial, Final, and Replace event, plus a completion sidecar for streamed file output. It is open and unmerged as of 2026-09-25.
• Keep this PR in draft for the user-facing native live transcript feature. After the upstream change is merged and released, add a reader for that mirror to Doubao Say's overlay, verify its lifecycle against the released binary, and then mark this PR ready. The compatibility fix in the current diff handles final text only; it does not depend on an unreleased Voxtype interface.

Real test

• Downloaded Whisper multilingual base and tiny without activating either as the system default. Used official Voxtype 1.1.0 AVX2 binary in an isolated daemon, verified against its release SHA256SUMS.
• Played a 5.59s Chinese fixture through a temporary PipeWire virtual source; restored the original source afterward.
• Before fix, Doubao Say rejected streaming. After fix, its real Voxtype client returned 開放時間早上酒點,至於下午5點。 with no error. The final result enters the existing polishing path.
• Voxtype native streaming produced multiple interim recognitions, but the file did not exist at +3s and status exposed only state. The current public CLI delivers file text at stop; this PR does not add a separate partial transcript feed.
• Whisper base exceeded the default 0.8s update interval at 2.4–3.0s per inference and timed out after stop. tiny at a 1.5s interval took about 1.1–1.2s per inference and completed.

Checks

• Focused Voxtype tests passed.
• make check PYTHON=/home/leyang/doubao-say/.venv/bin/python passed.
• make coverage PYTHON=/home/leyang/doubao-say/.venv/bin/python passed (61% total).
• Local nonvisual verification notes: artifacts/verification/voxtype-streaming-compat/verification.md (ignored by Git).