CI-greening loop has no circuit breaker when jobs can never start (billing-locked account); log download 404s raise GitHubError #116

Found while dogfooding the 24/7 dev-worker loop (developerz.ai) against a repo whose GitHub account was Actions-billing-locked — every CI job "fails" in ~2s with zero steps and no logs.

Two behaviors compound into an infinite session burn:

1. Log download of never-started jobs raises GitHubError. Jobs that never start (billing lock, also plausible for queued/in-progress) have no logs; gh returns HTTP 404 per job. claudetm logs Could not save CI failures: ... GitHubError each cycle instead of classifying "no logs available" as "job never ran".

2. No circuit breaker in the greening loop. Each CI check spawns a fresh Opus session which correctly concludes "PR is fine, CI failure is external, nothing to change" — then the loop re-checks CI and spawns another session, burning subscription quota until --max-sessions. Repeated identical conclusions ("no changes needed") + identical CI signature should trip a breaker: post one PR comment explaining CI is externally blocked and stop cleanly.

Repro signature: all jobs conclude failure in 2-3s, steps: [], check-run annotation "The job was not started because your account is locked due to a billing issue."

Observed with claudetm v0.1.48, claudetm start "<goal>" --no-auto-merge --max-sessions 10, PR dinndev2/adminer#25.