refactor: question-triggered skill, npx skills install path

[mattfogel]

Summary

The prov skill stops bloating agent context, and the duplicate install paths collapse into one mechanism per responsibility.

1. Skill is now question-triggered. It no longer runs prov log before edits, refactors, or debugging. It activates only when the user asks an origin/intent/history question — "why does this do X", "what was the prompt for this line", "what's the history of this file". The old shape pulled provenance into context on every substantive edit, even when the user never asked for it.

2. Install paths split by responsibility. Capture hooks live at agent-hooks/hooks.json and install via prov install --agent claude (unchanged behavior from a user's perspective). The skill moves to skills/prov/ and installs separately via npx skills add mattfogel/prov — Vercel's skills CLI auto-discovers skills/<name>/SKILL.md from the repo root, so no path suffix is needed. The Claude Code plugin manifest is gone, eliminating the /plugin install prov plus prov install --agent claude overlap that could register the same four hooks twice.

Skill content changes

• description: triggers on provenance questions only, with an explicit "Do NOT auto-run before edits" line.
• Dropped the paths: glob — triggering is the question, not the file.
• Body covers what prov is, three query shapes (point lookup, whole-file, prov search), and how to compose the answer (cite the prompt verbatim; surface drift state; handle no-provenance plainly).
• references/triggers.md rewritten from "is this edit substantive" to "what phrasings should fire this skill and what query maps to each", with a clear "should NOT trigger" section.
• Smoke-test plan adds two explicit negative scenarios — a refactor request and a greenfield write must NOT invoke prov log. Those are the regression this rewrite exists to prevent.

Install path changes

• plugin/ → agent-hooks/. plugin/hooks/hooks.json flattened to agent-hooks/hooks.json; the embedded include_str! path in install.rs follows.
• plugin/.claude-plugin/plugin.json deleted — no marketplace install surface anymore.
• plugin/skills/prov/ → skills/prov/ so npx skills add mattfogel/prov works with no path suffix.
• prov install --plugin flag and print_plugin_instructions() removed, along with the install_plugin_flag_does_not_touch_repo test that gated them.
• cli_plugin_layout.rs → cli_agent_hooks_layout.rs; the plugin-manifest test is gone, hooks-bundle and README tests point at the new flat path.
• READMEs reworked: top-level lists npx skills add mattfogel/prov as an optional follow-up to prov install --agent claude; the new agent-hooks/README.md describes just the hooks bundle and points at skills/ for the read surface.

Capture hooks unchanged

The four Claude Code hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop) write to <git-dir>/prov-staging/ only — nothing they emit reaches the agent's prompt — so they aren't part of the context-bloat problem this PR fixes. They keep firing exactly as before, and prov install --agent claude still merges them into .claude/settings.json from the same embedded source.

Verification

./scripts/check.sh passes (build, test, fmt, clippy). The skill content lints in cli_skill_layout.rs and the renamed cli_agent_hooks_layout.rs gate the load-bearing on-disk shape. The behavioral skill triggers (fires on "why does this do X", stays silent on "refactor this") are verified manually per skills/prov/tests/skill_smoke.md — there is no automated harness for trigger fidelity.

---