bench(lab): admit a repository with one command

[luuuc]

Problem

Admitting a repository to the bench is three manual acts today and nothing checks that they agree with each other: a human clones it, hand-writes lab/repos/<id>.json with a commit copied out of the clone, and runs sense scan over it.

Two things go wrong and only one of them is loud. A sha typed into the json that is not the sha the clone sits at produces a worktree at the wrong tree for every run afterwards, and nothing says so. A scan that indexed half the tree produces a repository that reads as dark when the tool simply did not finish.

Summary

sense-lab repo <handle|url|path> does the whole thing: resolve, clone, pin, index, record.

The resolver is pure and its order is first match wins — an admitted id, a path on disk, owner/name, then a url. A local directory named owner/name resolves as the path, because a directory that is there is evidence and a handle is a guess.

Nothing is written before the resolution is printed. Everything that decides runs first and touches nothing; the decision is announced; only then does anything reach a disk. The recorded pin is read back out of the clone rather than carried over from the request.

Ownership decides what may be moved: a clone the lab made and that drifted goes back to its pin, and a clone somebody handed in is read and never written to, with a mismatch refused and its reason given.

Re-running on an admitted repository admits nothing and prints where it stands, because the crank will call this on every invocation.

Changes

• lab/internal/repo: the pure resolver, the checkout (clone, adopt, correct or refuse) and the index phase, which runs sense scan and reads its counts through the MCP server in the checkout
• lab/internal/cli/repocmd.go: the command, its announcement, and the two records it writes — the index artifact and the repository file, the latter written last and only when there is an index to read its languages off
• catalog.Repo gains checkout, whose absence is the ownership record
• lab/README.md and lab/repos/README.md: the first step of the shortest useful path is now this command

Test plan

• make ci green: coverage floor with no new exception, zero complexity suppressions, lint clean
• new files hold 98.8–100% line coverage
• five mutations that a first draft of the tests let through are now killed, each verified by making the mutation and watching a named test fail: dropping -embed from the scan, asking the server outside the checkout, deleting the MCP handshake, selecting the reply without its id, and reusing an earlier reply's document for a call that failed
• the exchange is pinned against the real sense binary, not only against the stand-in
• end to end against a real repository: clone, pin, index, record, then a second run that admits nothing

[codecov]

Codecov Report

❌ Patch coverage is 99.37304% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
lab/internal/repo/checkout.go	98.07%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!