Preserve transcript text with optional malformed annotations

src/docs/src/AI/speech2txt.md

@@ -57,7 +57,7 @@ Fine-tune how transcription runs.
 **Gemini-specific options** (when `provider: 'gemini'`):
 
 - `language` (String): Language code (e.g. `en`) the audio is spoken in. The result echoes it as `language`. Without it, the result has no `language`.
-- `timestamp_granularities` (Array): `['word']` adds a `words` array with each word's `start` and `end` in seconds. Only `word` entries are supported. Other entries fail with `bad_request`.
+- `timestamp_granularities` (Array): `['word']` adds a `words` array with each word's `start` and `end` in seconds. Only `word` entries are supported. Other entries fail with `bad_request`. When Gemini returns a word with an unusable timing or speaker, that word is left out of `words`, and `text` still has the full transcript.
 - `diarize` (Boolean): When `true`, the result includes `words`. Words can include a `speaker` label such as `spk:0`.
 
 Gemini rejects options it can't honor with `bad_request` instead of ignoring them. These are `prompt`, `temperature`, `logprobs: true`, `stream: true`, `chunking_strategy`, `known_speaker_names`, `extra_body`, and the xAI-specific options.
@@ -66,6 +66,8 @@ Gemini accepts WAV, MP3, AIFF, AAC, OGG, FLAC, MPEG, M4A, L16, Opus, A-law, mu-l
 
 Before transcribing, Puter reserves the cost of the counted audio tokens plus the full 8,192 output tokens, then charges the tokens Gemini reports. A transcription Gemini answers is charged even when it fails, as Gemini bills it, including one cut off at the output limit. When Gemini reports no usable token counts, only the counted audio is charged.
 
+On a Puter server that doesn't have Gemini set up, Gemini transcription fails with `503 service_unavailable` and isn't charged. Test mode still works there.
+
 #### `testMode` (Boolean) (optional)
 
 When `true`, skips the live API call and returns a static sample transcript so you can develop without consuming credits.