Document or align skill-install vs Read/Grep severity gating

Why

The PostToolUse hooks in `src/lib/yara-hooks.ts` apply different severity gates to the same triaged match depending on which surface fired:

• Read/Grep: terminates only on `severity === 'critical' || action === 'block'`; medium/low matches warn-and-continue via `additionalContext`.
• Skill install: terminates on ANY post-triage match, with no severity check.

A medium-severity prompt-injection rule (e.g., `prompt_injection_chat_markup`) that warrants a warning on a project file becomes a full session abort on a skill download.

There's a code comment marking this as INTENTIONAL ASYMMETRY (skills are untrusted external code), but no test pins the contract and no doc explains the threat-model reasoning.

What needs to happen

Pick one:

Option A — Document and lock in. Add a test that pins the asymmetry: a medium-severity match on Read/Grep should NOT terminate; the same match on skill install should. Update ARCHITECTURE.md with the threat-model reasoning.

Option B — Align severity gates. Apply `severity === 'critical' || action === 'block'` to skill install too.

Option A is probably right — skills genuinely are higher-risk content. But the asymmetry should be a deliberate, tested decision.

Anchor

`src/lib/yara-hooks.ts` — the skill-install hook handler (search for "INTENTIONAL ASYMMETRY" comment).

🤖 Generated with Claude Code

Start in src/lib/yara-hooks.ts at the skill-install handler and the “INTENTIONAL ASYMMETRY” comment, then compare its gate with the Read/Grep PostToolUse path. Choose whether to document and lock in the asymmetry or align the gates; update ARCHITECTURE.md and add the corresponding regression test so the selected behavior is explicit and tested.