Module scars 

Real scars: analyzer input source that reads review / friction notes from the workspace ledger, and emits High-severity lessons only when the same problem shows up twice or more — havamal’s “no scar, no doctrine” rule made mechanical.

Why: SELECTOR2 opened the two natural High paths (supersede / conflict), but analyzer output is still generic self-help (“break the task down”). The ledger’s highest-judgment content — human-written review findings and friction notes — was never read. This module reads it, groups by problem family, and only proposes candidates when a family recurs.

Assumptions (queue 324 goal — operator can override at gate):

• Same problem = same tag family (review or friction) + normalized first-line prefix equal (case-fold, whitespace-collapse, strip trailing punctuation, prefix limit 80 chars).
• Recurrence window = 30 days from the most recent occurrence.
• Per-run cap stays at 3; real-scar lessons take priority slots ahead of statistical heuristic lessons.

Structs

LedgerNote

One ledger note fed into the scar analyzer.

Constants

PREFIX_LEN

Family-key length: short enough to catch “same problem, different wording”, long enough to distinguish unrelated tickets. Bumped down from 80 after a real friction pair collided at the semicolon (queue 324 test fixture). pub since q328 SIGNCHECK so sign_check reuses the exact same family key (single vocabulary, no rival contradiction concept).

RECURRENCE_WINDOW_DAYS

Functions

analyze_recurring_scars

Recurring-problem analyzer output — folded into the main postmortem lessons as High severity (single-family cap: one lesson per recurring family).

normalize_prefix

Case-fold, whitespace-collapse, cut at first clause break (;, —, full-stop or Chinese 。), strip trailing punctuation, then take up to PREFIX_LEN chars. Two writeups of the same scar rarely share a full sentence, but the clause-head is stable.

notes_from_events

Extract scar notes from workspace-ledger note events. One event carrying multiple scar tags fans out to multiple LedgerNotes (a friction that also carries review counts as both — that’s their real classification, we don’t second-guess the writer).

read_ledger_notes

Read notes from a edda log --json-style JSONL file (row-per-event). Malformed lines skipped; missing file ⇒ empty. Kept as a testable path alongside notes_from_events (which reads structured Events directly).