Polish one detection rule

Polish one detection rule

Goal: improve one existing rule on both axes — catch real-world variants it currently misses (recall), and stop firing on benign content it should never have flagged (precision) — without losing true positives. One rule, one PR.

Every cycle starts at step 3 — what the rule does to the real corpus today — before touching anything. A rule can carry false positives from the day it shipped: it was written against synthetic fixtures, and the corpus has only ever been checked for regressions, never audited.

Read docs/rule-verification.md for the detection approach and confidence math behind each rule. (Pack layout and the RE2 constraint are in CLAUDE.md.)

Guardrails

Generated attack payloads are inert test data only — never execute them. All the global guardrails from sg-maintain apply.

1. Pick the rule

Load state from .claude/maintenance/state.json → rule_last_polished. List the rule IDs across the packs:

grep -h 'id: SG-' pkg/rules/packs/*.yaml

Choose the rule with the oldest (or missing) rule_last_polished timestamp. If a caller named a specific rule, use that instead.

2. Understand its current coverage

• Open the rule's YAML block: its match tree, confidence, suppress list, targets.
• Read its section in docs/rule-verification.md (Signals / FP carve-outs / Fixtures).
• Note what it already matches and, importantly, its documented false-positive carve-outs — new test cases must respect these, and step 3 will tell you whether they actually hold in the wild.
• Check whether a previous cycle already recorded an FP audit for this rule (the "Corpus precision" note step 8 writes). If so, that is your baseline: look for what changed, not the same ground.

3. Audit its real false positives on the corpus