[ygivenx]

FreeWispr can crash or retain inconsistent recording/transcription state when Teams changes the default microphone or a dock carrying the Yeti and monitors disconnects. This consolidates the relevant fixes from #3 and #13 on top of the configuration-change recovery already merged in #6, and prepares version 1.3.2.

Changes

• Open a fresh microphone engine for each recording, use its live format, retry startup once, and convert native Objective-C audio exceptions into recoverable errors.
• Restore the busy-microphone guard: while another app uses the default input, show a temporary busy message. Retry after that app releases the mic.
• Deliver route-change callbacks on the main queue, ignore previous sessions, discard interrupted audio, and reset the recording UI.
• Isolate recorder, transcriber, and indicator state to the main actor; reserve the pipeline before scheduling transcription and prevent overlapping inference/model switches.
• Clean up configuration observers and hotkey event taps; reposition the indicator after screen changes; start setup from the app lifecycle.
• Feed each buffer to the converter once. Add hardware-free regression tests, a native exception fixture, and a PR unit-test workflow with coverage.
• Notarize/staple/verify the DMG as well as the app in both release workflows, run unit tests before packaging, and assess Gatekeeper only after notarization in the local script.
• Validate stable release tags against VERSION and Info.plist; correct the arm64-only download requirements and add a release/distribution guide.
• Bump VERSION and both bundle version fields to 1.3.2 and update the changelog/architecture notes. Correct the existing model-validation test fixture to use the little-endian GGML header already required by production code.

Validation

• Debug build and optimized arm64 release build passed.
• GitHub unit suite with coverage passed: 53 tests, zero failures, including all eight microphone regression cases (repeated failure recovery, busy-mic recovery, interruption/stale-session handling, format changes, late callbacks, and native exception recovery).
• All eight microphone cases also passed locally using a standalone runner; full local XCTest execution is unavailable because the installed command-line tools lack XCTest.

Release integration audit

• Downloaded v1.3.1: SHA-256 matched GitHub, app signature validated, Gatekeeper accepted the app as Notarized Developer ID, and the artifact's signing certificate expires February 1, 2027.
• The v1.3.1 DMG itself was rejected as Unnotarized Developer ID despite a valid signature. The new DMG notarization/assessment steps must be proven with a signed testing release before stable publication.
• All eight Hugging Face model/encoder URLs returned HTTP 200 for HEAD requests.
• Workflow YAML, every workflow shell block, and the local packaging script passed syntax checks; version validation accepted v1.3.2 and rejected mismatched/prerelease tags.
• Required signing/notarization secret names are configured. Gumroad product/content/customer-flow verification awaits the product URL; no Gumroad repository integration was found.

Hardware verification

1. With the Yeti connected through the dock and active in Teams, trigger FreeWispr: it should show a busy message and remain running.
2. End the call and retry dictation after Teams releases the microphone.
3. Disconnect/reconnect the dock during dictation: the recording should stop, the indicator should clear, and incomplete text should not be pasted.
4. Repeat hotkey presses and model-selection attempts while transcription is active.

The exact Teams/Yeti/dock setup has not been reproduced locally. Related: #3, #6, #13; the earlier PRs remain open for separate review.