Delete repetition/thrash/backstop detection outright

[TheGreatAxios]

Summary

Deletion task (CL-6995): removes all repetition/cycle/no-progress-by-turn-count detection outright, no replacement, no raised thresholds.

Deleted entirely:

• src/subagent/repetition.ts (detectRepetition, trackContentlessGrowth, detectTailCharLoop)
• src/subagent/repetition.test.ts
• src/util/period-detection.ts (detectSequencePeriod) and its test — nothing outside the deleted detectors imported it
• src/subagent/stop-policy.test.ts — tested only detectToolFingerprintThrash / detectTurnsSinceUserMessageBackstop, both deleted

Removed from existing files:

• src/subagent/run.ts: the streamed-text repetition/contentless-growth abort path in the sub-agent stream sink, repetitionStopDetail/contentlessGrowthDetail
• src/subagent/stop-policy.ts: detectToolFingerprintThrash, IDENTICAL_REPEAT_MIN/CYCLE_REPEAT_MIN, TURNS_SINCE_USER_MESSAGE_BACKSTOP, MAX_LEAF_PROGRESS_BACKSTOP_RESETS, the "repetition" ForcedStopReason and its parent-hint text, "salvage-repetition" outcome
• src/agent/director.ts: the tool-fingerprint thrash pause and turns-since-user-message backstop nudge/escalation (pausedForToolOnly, toolFingerprintHistory, backstopNudgeFiredAtTurn, leafProgressBackstopResets, etc.) — kept the unrelated soft check-in nudge (toolOnlyTurnNudgeAt), which is a plain turn-count nudge, not a comparison of model output
• src/subagent/brief-dispatch.ts: "repetition" hard-block salvage class
• src/tui/turn-state.ts: the intra-cycle detectTailCharLoop check and its throttling bookkeeping (streamCharsSeen, repetitionCheckedAt) — kept the independent cross-cycle fingerprint streak (consecutiveMatchingCycles), which is unrelated TUI stall-recovery machinery not named in scope
• src/session/summarizer.ts: dropped a detectRepetition call gating which assistant turns feed the compaction summary excerpt (found as an additional call site not in the original inventory)
• src/session/stream-journal.ts: "repetition" from PartialFlushReason

Dead-by-consequence cleanup: unused imports/helpers left over from the above (fingerprintToolCalls/backstop imports in director.ts, several now-unused test fixtures in director.test.ts, stop-policy.ts's now-unused detectSequencePeriod import).

Tests: deleted tests that asserted the deleted detectors fire (turn-monitor, turn-state, stall-watchdog, director, index, summarizer suites); added one test in src/subagent/index.test.ts showing a leaf that keeps varying its own tool calls in reaction to a stably-failing environment runs to the turn budget and still completes normally on a real report.

Where the line is drawn

Kept as-is (transport reliability, not behavior policing):

• Provider stream error / connection failure handling and inference retry/backoff (retry-policy.ts, the harness's own attempt budget)
• Turn budget (maxTurns) and no-progress (identical tool-call fingerprint repeated DEFAULT_SUBAGENT_REPEAT_LIMIT times) on leaf sub-agents — this is a simple repeat-count check, not a comparison/hash of model output, and neither peer's absence-of-repetition-detection claim covers turn caps (both peers were found to have no default turn cap at all, which is a separate, unaddressed gap noted for future consideration, not touched here)
• The independent cross-cycle TUI stall-recovery fingerprint streak in turn-state.ts (not named in the deletion scope and not part of the repetition.ts/stop-policy.ts machinery)

Test plan

• bun run check green in the worktree (lint, typecheck, build, test — 5366 pass / 0 fail)
• scripts/verify-corbits-only-scope.sh passes (no vendor/ changes)

Net: -2620 lines (118 insertions, 2738 deletions across 24 files).

[linear-code]

CL-6995

[TheGreatAxios]

Superseded by #612, which re-applied this deletion on a settled main after #610 restructured the same files. #612 merged; the work is in 0.3.0.