fix(ttsr): interrupt paragraph loops by near-duplicate frequency

[code-yeongyu]

Summary

Closes the near-repeat gap #1330 describes, with a frequency rule instead of a looser match.

paragraph-repeat (#1326) compares paragraphs byte for byte over a 64-entry ring, so it misses two shapes. A model that restates the same step in different words every time never produces three identical hashes. And a cycle longer than the ring stays invisible even where it repeats verbatim.

A captured incident hit both at once: claude-fable-5-1 streamed narration for 12 minutes with zero tool calls, paraphrasing one action ("I'm assembling the final delivery ... — task complete." / "Both batches sent with photos and reaction cleared — done."), then settled into an exact 114-paragraph cycle. Replaying that text through the shipped collapseDetector returns no match on any mechanism. Nothing ended the generation; the user killed the session.

What this adds

detectors/collapse-near-duplicates.ts (near-duplicate-paragraphs), wired last in the collapse chain for text and thinking streams:

• a paragraph is an echo when its normalized word set reaches 0.5 Jaccard against any of the last 32 eligible paragraphs (eligibility is the existing 64 chars / 24 word chars);
• the detector fires only when at least 8 of the last 12 eligible paragraphs are echoes.

The density gate is what keeps ordinary prose out. #1330 warns that a looser match alone would interrupt a legitimate "processing item N" traversal; a rate gate answers that directly, because one similar pair, a callback to an earlier point, or a closing summary never reaches 8-of-12. Cycle length does not matter, so the 114-paragraph case is covered too. Paragraphs inside fenced code blocks are skipped and never enter the history.

Remediation reuses the existing collapse path: abort, truncate from the first echoed paragraph in the firing window (its anchor is kept), then the collapse nudge.

Calibration evidence

Replaying the shipped detector over the local session store - 12,499 real assistant text and thinking parts, 28.9 MB - fires twice, and both firings are known runaway generations from one 2026-09-03 session (38,193 chars, and 363,042 chars ending in stopReason: length). No other match.

input	result
incident tail (sanitized fixture)	fires after 1,935 chars
12,499 real assistant parts	2 matches, both known loops
distinct multi-sentence prose, 40 paragraphs	silent
10 echoes spread across 40 paragraphs	silent
64 KB healthy prefix	silent
24 near-identical fenced code blocks	silent

Tests

test/ttsr/detector-collapse-near-duplicates.test.ts replays the sanitized incident tail, pins chunk-boundary independence, and pins every negative above plus tool-stream exclusion and the hand-off to paragraph-repeat on byte-identical cycles.

Mutation proof: removing the one wiring line in collapse.ts fails exactly the three positive tests (incident replay, synthetic paraphrases, thinking stream); restoring it returns 9/9.

Fixture changes, and why they are not loosened assertions

Two synthetic generators varied only by a counter, so after digit normalization every paragraph carried the same tokens - healthy input for a byte-exact rule, a narration loop for a normalized one:

• collapse-test-inputs.ts buildHealthyPrefix now composes varied vocabulary;
• detector-collapse-paragraphs.test.ts narration() now emits seven lexically distinct steps.

Their assertions are untouched, so the exact-repeat contract still tests exact repetition. The digit-only-drift shape those fixtures used to have is the shape #1330 asks to catch, and it is now covered by the new test.

Real-CLI QA

A new senpi-qa mock-loop scenario, ttsr-near-duplicate-loop, streams fifteen paraphrases of one action (none byte-identical) from the local fake model server through the real CLI:

[PASS] CLI exits zero - code=0
[PASS] exactly two model turns (abort + one bounded recovery) - requests=2
[PASS] no paragraph repeats byte-exactly, so only the frequency rule can fire - distinct=15/15
[PASS] persisted aborted message is truncated before the streamed tail - stopReason=aborted chars=527/2446
[PASS] recovery request never replays the streamed tail - tailPresent=false
[PASS] collapse-repetition system-interrupt injected into the recovery request
[PASS] recovery answer returned
7/7 passed

The existing scenarios still pass on the same build: ttsr-paragraph-loop 6/6, ttsr-collapse 4/4.

Verification

Run on a separate macOS machine (bun 1.4.2) against origin/main 9e1f97f634:

• bunx vitest run test/ttsr - 18 files, 321 tests passed
• bun run check - exit 0
• bun scripts/check-pr-changelog.mjs - PASS

Evidence for the incident itself is held locally; the in-tree fixture is the sanitized tail.

Fixes #1330