bench(reviewer-eval): weekly mining routine — local cron sweep, notify-only (sc-1415) - #321

…ock exit propagation with a test (sc-2492) (#568)

Both logged sweeps (2026-08-24, 2026-08-31) failed every stage with nothing in the log but bun's "ulimit -n 2147483646 / launchctl limit maxfiles" hint. Two things were wrong, both in the runner.

- The 6-line tail cut off the error line printed just above that hint, so the cause of two failed weeks is unrecoverable. run_stage now keeps each stage's full output under ~/.claude-usage/weekly-mining/<sweep>.<unique>/<stage>.log (pruned by age after 8 weeks, never by count, so a concurrent sweep's directory is never touched) and, when a stage fails, logs its last 20 lines instead of 6. bun's root-error output is the last thing the process prints (error line, blank, 8-line hint block), so a 20-line failure tail always contains it; the full file is there for anything longer. No classifier decides what an "error line" is.
- Under cron the soft fd limit is 256 (launchctl maxfiles 256 unlimited), and bun's crash handler prints that same hint for ANY root error while the soft limit is under 16384 (src/crash_handler/lib.rs, the `Unexpected` branch), not only for EMFILE. A low limit therefore hides whatever the real error was. The runner lifts the soft limit to the hard limit (capped at 65536) before any miner runs, and the sweep header records the effective limits.

Is the exhaustion inherent? Every miner is synchronous (execFileSync only, no fan-out) and a cron-like replica (env -i, ulimit -n 256, the script's own PATH) ran mine-ghsa and mine-telemetry to exit 0; mine-bots was still sweeping 597 frink PRs at ship time and propose-bots depends on its output, so those two results are posted on sc-2492 rather than claimed here. No descriptor leak was found in the miners, so raising the limit is a diagnosability measure that also removes the one way a 256 limit could fail a legitimately larger sweep.

Exit propagation already existed (finding on #321); the regression test now locks it: a failing stub miner fails the sweep with exit 1, names the stage with its full-output path, puts the error line in the log, keeps the full output, and skips propose-telemetry; all-stages-failing names all four; two sweeps in the same minute get separate directories; a clean sweep exits 0. Stubs go through the script's own PATH ($HOME/.bun/bin first) with osascript stubbed so no notification fires.