plugin pipe: degenerate-turn one-shot retry is dead on the Gemini wire — gate eligibility + per-lane answer-accounting asymmetries (checklist) #2769

来源: #2764 (fix #2756) 分析退化轮门控 (answer-text gate) 时发现

Context

PR #2764 redefines the plugin pipes' one-shot degenerate-turn retry (#732/#821) gate as "did the host receive ANSWER text?" (an answerProse accumulator alongside byte-exact visibleTextChars). While reviewing it, three unverified facets of this SAME mechanism family surfaced (per-wire gate eligibility + answer accounting, one failure domain), managed here as one checklist.

Checklist

1. The gate is unreachable on the Gemini wire (the retry machinery is dead there; pre-existing)

• The turnTerminal assignment at the chat-pipe call site (src/plugin.ts, ~line 2915 at fix(#2756): degenerate-turn gate asks for ANSWER text, not visible bytes #2764 head) covers anthropic (message_delta) and openai (finish_reason) ONLY — there is no google branch, so retryEmptyTurn() is never invoked for protocol "google".
• Even with the branch added: Gemini's terminal reason is UPPERCASE "STOP" while CLEAN_TURN_REASONS = {"stop","end_turn","stop_sequence"} — the reason check would refuse every retry anyway.
• Impact: the [BUG] In some cases, with Codex's auto model + DeepSeek, it will suddenly terminate automatically. #732/workbuddy报错Empty stream: upstream gateway sent only placeholder chunks without any model output (chunks=0, bytes=0) #821 stall protection (and the DSH native mode: assistant text ending with </summary> is treated as a completed step, so the turn silently stops (149 samples from session transcripts) #2303 draft-tail retry, which lives inside the same function) simply does not exist on the Gemini wire. Pre-existing since those mechanisms landed; NOT introduced by fix(#2756): degenerate-turn gate asks for ANSWER text, not visible bytes #2764.

2. Google-lane answer-accounting asymmetries (must close BEFORE activating item 1)

• Fast path: processGoogle's fast path incremented visibleTextChars without feeding answerProse — fixed in fix(#2756): degenerate-turn gate asks for ANSWER text, not visible bytes #2764 review commit 392c7a1 plus a regression test. Left unfixed, the moment the gate reaches this lane EVERY healthy thought+prose turn would read as zero-answer and fire a spurious one-shot retry (double-billed turn, duplicated continuation).
• Fold site (processGoogle, ~line 2782): unlike the openai/anthropic lanes it never consumes droppedTagInFrame. Today the consequence is classification-only (post-drop released bytes count into answerProse instead of releasedMarkupChars; net gate outcome for affected turns matches master because sawStrippedEcho covers the same turns), but the lane should reach parity in the same change that activates item 1.

3. OpenAI-fold slice-subtraction corner (#2764 new code, latent)

At the openai fold site, when droppedTagInFrame && tailLen > 0 && the folded tail is non-orphan && |tail| < |released|, answerProse = answerProse.slice(0, answerProse.length - released.length) removes bytes from the TAIL instead of from released (keeps a prefix of the released markup, drops real tail prose). Reachability needs an exotic combination (e.g. an absorb-instructed tcHold flush co-existing with a dropped tag in the same terminal frame); TagEchoFilter.flush() CAN return non-orphan prose, so the corner is reachable in principle. Fix: pin flush()'s return semantics with a targeted filter test first, then replace the subtraction with explicit branch accounting (exclude released when dropped-tag, include the tail when non-orphan).

Suggested path

1. Wire item 1: add the google turnTerminal branch AND accept "STOP" (case-normalize or extend the set), keeping the clean-stop semantics identical to the other wires.
2. In the SAME change, complete item 2 parity (google fold-site droppedTagInFrame consumption).
3. Prove both directions with tests: a gemini degenerate turn (thinking-only or whitespace-only content) retries exactly once; a healthy gemini turn (thought part + real prose) never retries and logs no refusal.
4. Item 3 separately, after the filter-semantics pin.