On-device dictation can ignore model startup errors and report an empty completion

Does this issue occur when all extensions are disabled?: Not tested; reproduced with controlled dependencies executing the source methods.

- VS Code Version: Startup gap verified in 1.137.0 and 1.138.0 source; reproduction at `4fa9fcf245ad90eb1f21fae5ed0511515e068b1d`.
- OS Version: Windows; source-method reproduction, not a real-microphone test.

Steps to Reproduce:

1. Start local dictation with a test backend that fails immediately and returns model status `Error` from `getModelStatus()`.
2. Stop dictation after the renderer enters Recording.

**Actual:** No startup failure is surfaced; stop returns no text and records `completed` with `transcriptLength=0`.

**Expected:** Surface the startup error and terminate the session as failed.

### Potential root cause
`_startLocalSession()` skips status tracking when the initial snapshot is already `Error`. `_failSession()` also ignores failures while the renderer is still Idle during startup. Native `stop()` then catches the failed-open promise and returns empty text, assuming the error was already handled.

Sources: [renderer startup](https://github.com/microsoft/vscode/blob/4fa9fcf245ad90eb1f21fae5ed0511515e068b1d/src/vs/workbench/contrib/chat/browser/speechToText/chatSpeechToTextService.ts#L1045-L1078), [startup failure guard](https://github.com/microsoft/vscode/blob/4fa9fcf245ad90eb1f21fae5ed0511515e068b1d/src/vs/workbench/contrib/chat/browser/speechToText/chatSpeechToTextService.ts#L1250-L1265), [native stop](https://github.com/microsoft/vscode/blob/4fa9fcf245ad90eb1f21fae5ed0511515e068b1d/src/vs/platform/localTranscription/node/localTranscriptionService.ts#L715-L743).

The combined startup/stop path was reproduced with synthetic dependencies; its production frequency is not established.