Learn from Sage, detection-gap (miss) analysis (V2, file-centric)

Learning is not summarizing what Sage found. It is finding issues that leaked past review and shipped, then working backwards to close the gap so the reviewer catches the next one.

Two files, one rule each:

• learned-patterns.md, the canonical, consolidated ruleset. The only file a review loads as heuristics. Human-editable.
• learned-patterns.candidate.md, staging for new learnings. Staging an exact equivalent (same title, scope, impact and guidance) adds <!-- seen:N --> and <!-- last:<date> --> to the existing entry instead of appending; anything else is appended. Reviews do NOT read it. It is merged into learned-patterns.md only when a human triggers consolidation (a one-shot AI merge), after which it's cleared.

The interpreter in every command below

Commands here are written as <python> …. Replace <python> with the absolute interpreter path the task prompt names, it is the one the app itself runs, and it is the only interpreter guaranteed to exist on this host. Never substitute a bare python3: Windows has no such interpreter (the name resolves to a Microsoft Store app-execution alias that runs nothing), so the command would stage nothing and the learning would be silently lost. Outside a review session, use the interpreter running the app.

Where every path below is rooted