Function run_train_ci 

pub fn run_train_ci(
    repo: &Path,
    tip: &str,
    spec: &CommandSpec,
    ci: &mut dyn CiExecutor,
) -> Result<Verdict, String>

Checks one train commit with a CiExecutor instead of asking the host forge (D18).

The forge path and this one answer the same question by opposite means. Asking the forge requires publishing the speculative merge of every open PR to a remote and trusting whatever ran there; this checks the train in the worktree that just built it, with an operator-declared command, on a provider the operator picked. What comes back is a Verdict rather than a bool, so the caller can still tell “the train is bad” from “we could not find out” — the distinction the whole seam exists for, and one the forge’s status/conclusion enums also make and the bool did not.

repo is checked out at tip first, forcibly: the train build owns this worktree and left it detached at the train it built, but an advisory differential run between the two may have moved it.

The job is content-addressed by tip under git’s own codec, so the subject a cache keys on and the subject a RecordCheck names are the commit itself rather than a queue-local id. may_write_cache stays false: a train holds unreviewed code from every open PR, which is exactly the untrusted case the flag is for.

Errors

Git could not resolve or check out tip, the executor could not be reached, or it answered with something other than one verdict for the one job. All three mean we did not find out, never that the train is bad — a caller must not land on this and must not blame a change for it.