docs+feat: tighten the standards reference, and deliver review conventions

[yuema137]

Two small changes, kept small on purpose.

1. The standards reference, 175 lines to 136

Cut explanation that repeated itself, not anything it discloses. Every limit stays: an absent tool is INCONCLUSIVE, approve does not enforce who approves, trigger: "pr" has no hook, and reading a file proves nothing about whether the agent followed it.

2. Review conventions reach the agent

The last item in the plan, in the smallest honest form. Whatever a project set to the commit trigger rides along in one report: the checks run, and the declared conventions are restated for the agent to apply to what it just committed.

This needed no new delivery path — the hook already sends a message at that moment, so the conventions are about ten lines of content added to it rather than a mechanism of their own. That is what the earlier plan would have built, had the Stop handler survived.

Conventions with the pr trigger are not orphaned: SKILL.md already routes the agent to the standards contract whenever a project declares one. The hook adds a commit-time reminder; it is not the only path, and the documentation says so.

Not done, and why

I proposed removing Step 3's approval machinery and then retracted it, so it is untouched here.

The claim behind that proposal was that it "does not enforce who approves, which its shape implies". Re-checking against the threat it exists for, that was wrong. The realistic threat is a contributor's pull request adding or changing a command that then runs when you check the branch out. The approval is bound to the command digest and lives outside the working tree, so a fresh clone has none and any change to a command revokes it — that threat is covered. What it does not cover is a local agent running approve, which is the same class as an agent running the installer, and not what the boundary was for.

It is also 72 lines out of 752, not the heaviest thing in the file. Deleting working code that blocks a real threat, on a premise I no longer believe, would have been the worse mistake.

The genuinely over-built piece was the Stop mapping, and that was already cut in #18.

Validation

Ran 165 tests ... OK (skipped=1); --check PASS on all five checks.

🤖 Generated with Claude Code

https://claude.ai/code/session_016cCs5DdUjLDkP8V3Z5x6XH