compress: break identical-rejection retry death-loops - #103

Incident

pi session 01a02542 (qwen3.8-27b via local vLLM/SGLang, billion-context-pi 0.1.43 + acp-kernel 0.0.30) entered a compress retry death-loop:

• After 5 successful compressions nothing compressible remained (22.6K context, 11 active blocks, remaining compressible content 0 chars < min 5000).
• The min-gate rejection ended with: "Nothing to do. Current active blocks span b1..b11 — retry with startId/endId set to active block IDs in that span."
• The model re-submitted the identical rejected call () 3853 times over 5h11m (~4.9s/call, ~800 GPU-minutes burned) — it followed the literal "retry" instruction and ignored the "Nothing to do" half. No dedup or circuit breaker existed anywhere.

Fix

1. Rejection tracking — CompressionState.rejections (new optional field; bounded 8-entry FIFO of { spec, count }; keyed on the range signature startRef..endRef list; backward compatible — old persisted states start counting from 1, cloned by cloneState).
2. Terminal message from the 2nd identical rejection — the gate responds: identical range(s) rejected N times, re-submitting will keep failing, stop calling compress with these refs and answer the user; new content becomes compressible once it exceeds min chars. No retry guidance of any kind.
3. First-rejection advice kept, de-weaponized — the liveHint now reads "only a block-to-block distillation (startId/endId set to those block IDs) can reclaim more; otherwise nothing remains to compress" — the word "retry" is gone, the genuinely useful tier-2 action is still offered.

Tests

• New tests/compress-rejection-loop.test.ts (4 cases): first rejection keeps guidance + records count; identical 2nd/3rd rejections flip to terminal (asserts the message contains "rejected N times" and contains NO retry/re-issue/acp_status guidance); different range gets its own counter; caller state not mutated; legacy state without the field counts from 1.
• Updated the one assertion that pinned the old "retry with startId/endId" wording.

Full pre-flight: tsc --noEmit clean, 395 tests pass, build OK.

Blast radius

Pure kernel gate change — both adapters (billion-context-pi, billion-context-omp) pick it up on their next kernel pin bump. The omp adapter's compress tool needs no change (it forwards kernel errors verbatim).