Motivation

• Implicit feedback parsing was too permissive and could infer tiers from descriptive text, producing false-positive corrections.
• A single corrupted JSONL line could break persisted reads and disable downstream summary/review flows.
• Duplicate feedback records were inflating calibration metrics and creating redundant open review items.

Description

• Tightened implicit-tier extraction in parseImplicitFeedback to only infer T1..T4 when correction-oriented cues are present (change in src/routing/feedback-loop.ts).
• Added normalizeReactionValue() and used it in fingerprintFeedback and reaction handling to ensure consistent reaction normalization (in src/routing/feedback-loop.ts).
• Hardened readJsonl() to skip malformed JSONL lines with a warning and continue returning well-formed records (in src/routing/feedback-loop.ts).
• Prevented duplicate feedback from creating review queue items and excluded duplicates from mismatch/false-escalation/latency/source aggregation while still counting duplicates (in src/routing/feedback-loop.ts).
• Updated and extended tests in src/routing/feedback-loop.test.ts to cover descriptive-tier negatives, corrupted JSONL tolerance, and deduplicated summary expectations, and fixed formatting drift in src/gateway/hitl-request-store.ts.

Testing

• Ran pnpm format and pnpm format:fix to apply repository formatting (succeeded).
• Ran pnpm test src/routing/feedback-loop.test.ts and the file's test suite passed (7 tests passed).
• Ran pnpm check, which reports pre-existing unrelated TypeScript/module issues in other areas of the repo that are not caused by this change.

Codex Task