v0.2.0-beta.5.1.8 - Gemini TTS Content-Blocked Recovery

Gemini TTS content_blocked classifier recovery

Trigger:

• controlled Batch A run 31587048578
• source v0.2.0-beta.5.1.7 / 30c75a0
• Atomic Habits Sulafat chunks succeeded; Schedar chunk 1 then returned HTTP 400 content_blocked

Fix:

• explicit speech-synthesis preamble
• spoken transcript START/END boundaries
• transcript preserved exactly
• no safety setting changes
• recovery only for exact transcript SHA already synthesized successfully in the same process
• every actual recovery-prompt POST consumes the global recovery budget
• generic retryable recovery responses also consume recovery budget
• max recovery POSTs = 2 globally
• third recovery POST refused before network fetch
• overall generation POST hard cap remains 10

Deterministic proof:

• primary content_blocked -> recovery POST #1 -> retryable 500
• recovery POST #2 -> retryable 500
• later eligible recovery POST #3 is refused before network fetch
• unseen transcript fails closed
• completed/no-audio materialization tests remain PASS

Lifecycle:

• Issue #68
• PR #69
• final PR head: f3f9d42
• merge/release SHA: 87da695
• PR CI: 31594232501
• main CI: 31594649547
• Pages deploy: 31594649570

Safety:

• no Batch A dispatch in this release step
• no media-dual-v0.2.0-beta.6 creation
• no episodes.json/player/Piper promotion