Deploy #449's core (rust-3) job failed in 5 seconds with "The runner could not fetch the job: Failed to read JSON: response body closed before all bytes were read": the runner asked for its spec once and gave the whole job up on one truncated answer. Api::spec now asks again, up to 4 times with 1, 2 and 4 seconds between, when the connection drops, the body ends early or g1t answers 5xx; never on a 4xx, which says no. Not with rotate (self-hosted runners): g1t spends the token on the first ask whether or not the answer arrived, so a second would be refused, as plan.rs's a_job_token_fetches_its_spec_once means it to be. g1t's own sandboxes take the runner from the image Deploy builds, so this reaches them with the next deploy.
Checked: cargo test -p g1t-runner (262, new: a spec cut short is asked for again and arrives; one that spends its token is asked for once; only what may arrive next time is retried); cargo clippy -p g1t-runner --all-targets. G1t-Session: ses_01m5bgsckce7gt943v64h219yy