Update gpt-6-astra-high: score the second SWE-Live sample (SWE-Live 10/19, overall 45/99)
Follow-up to #18 (merged). The 19 SWE-Live tasks of gpt-6-astra-high/ were run a second time on 2026-10-20 with the same configuration, harness (upstream main 5933dd2) and cached calibration gates: a main batch, one rerun of invalid / infrastructure_error / timed-out tasks, and the usual infra fix-up of swelive/0262 (its official reference patch failed native preflight, the README-known flake). That second sample (10 pass / 8 fail / 1 invalid) replaces the first (13 / 5 / 1) as the later attempt, so the directory now scores 45 pass, 51 fail, 1 invalid, 2 infrastructure_error (valid pass/fail 96/99; was 48 / 48 / 1 / 2). The separately labeled format-tolerant view moves from 50 to 47 pass accordingly.
Changed: all nine files of gpt-6-astra-high/. attempts.jsonl now lists 148 attempts (was 122): the first SWE-Live sample stays, marked scored false, and the second carries rounds "swelive-rerun", "swelive-rerun-rerun" and "swelive-rerun-fix"; task-artifacts.tar.gz adds those batches (10,258 members); notes.txt and manifest.json record the replacement. Same redaction and exclusions as #18; validation (99 unique task IDs, every scored attempt in the archive, member hashes, sensitive-pattern scans) passed locally.
Superseded by #20, which moves the Amazon Bedrock runs to amazon-bedrock// and includes this update (amazon-bedrock/gpt-6-astra/high/).