Update Sherlog skill for agent-first recall loop

[catoncat]

Parent

#86

What to build

Update the distributable Sherlog skill and current docs to teach the agent-first recall loop that now exists: use find, inspect the recall packet, run read-*, answer from evidence, and self-evaluate recall quality. Keep dogfood capture private and explicit.

Acceptance criteria

• The distributable skill describes the updated find -> recall packet -> read-* -> answer flow.
• Skill guidance explains elision metadata, zero-result refine hints, and when to sync versus when to refine.
• Skill guidance preserves dogfood boundaries: ordinary agents may suggest candidate capture but must not create/promote goldens.
• Current docs describe implemented behavior only, not aspirational architecture.
• Smoke checks verify the installed/global skill path and CLI help remain aligned with the published behavior.

Blocked by

• Deterministic large-message elision for read commands #89
• Minimal recall packet in find JSON #90
• Zero-result refine hints and coverage truth #91