fix: do not persist a deleted speech model as the active selection

[postoso]

Description

Deleting a non-active speech model used to briefly persist the model being deleted as selectedSpeechModel, then restore the real selection through an un-awaited async restore. If the app quit, crashed, or relaunched during that window, the deleted model could become the saved active model.

This routes per-model deletion through an awaitable ASRService.clearModelCache(for:) call, which clears the target model without changing the saved selection. It also refreshes installed-model state after deletion so the model cards update without relying on a temporary selection change.

The removed suppressSpeechProviderSync and skipNextSpeechModelSync flags were only needed for the old selection bounce. With the bounce removed, the speech provider and preview state can stay directly synced to the actual selected model. The view-model method is now awaitable, so the regression test does not rely on a fixed sleep.

Type of Change

• 🐞 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality not to work as expected)
• 📝 Documentation update

Related Issue or Discussion

Related to #467. This fixes the delete-non-active-model persistence path only; it does not claim to close every possible path that could reset the active speech model.

Testing

• Tested on Intel Mac
• Tested on Apple Silicon Mac
• Tested on macOS 15.7.7
• Ran linter locally: swiftlint --strict --config .swiftlint.yml Sources Tests/FluidDictationIntegrationTests/DictationE2ETests.swift
• Ran formatter locally: swiftformat --config .swiftformat Sources
• Built locally: xcodebuild build-for-testing -project Fluid.xcodeproj -scheme Fluid -destination 'platform=macOS,arch=arm64' CODE_SIGN_IDENTITY='' CODE_SIGNING_REQUIRED=NO CODE_SIGNING_ALLOWED=NO

Added DictationE2ETests.testDeleteNonActiveSpeechModel_doesNotBounceSelectionThroughDeletedModel. The test records every write to SelectedSpeechModel and fails if deleting .appleSpeech while .whisperTiny is active ever persists the deleted value, even transiently.

The test intentionally uses a built-in no-cache model so it can catch the selection-bounce regression without deleting a real downloaded speech model from the developer's machine.

Notes

This is intentionally narrow. It removes one concrete race in model deletion without changing model activation, update behavior, or broader speech-model fallback logic.

Full hosted test execution was not run locally because the installed FluidVoice app was open; the hosted test runner can hang in that state. build-for-testing compiled the app and test target, and CI should run the full hosted suite after the branch update.

Screenshots / Video

• No UI/visual changes; screenshots/video are not applicable.

[greptile-apps]

Greptile Summary

The PR removes the temporary selected-model swap previously used for cache deletion, directly clears the requested model cache, and refreshes installed-model state afterward.

• Makes speech-model deletion awaitable from the settings UI.
• Keeps provider and preview state synchronized directly with the actual selected model.
• Adds regression coverage that records persisted model-selection writes during non-active-model deletion.

Confidence Score: 5/5

The PR appears safe to merge with no actionable changed-code defects identified.

Targeted cache deletion preserves the selected-model cleanup lifecycle while avoiding the transient persisted-selection mutation, and the new test directly covers that regression.

_{Reviews (1): Last reviewed commit: "fix: do not persist a deleted speech mod..." | Re-trigger Greptile}