A cell that errors on a later attempt still discards the earlier attempts' transcripts #97

#94 keeps the transcript of every superseded attempt, but only on the paths where runOne returns. Two guards inside the loop throw: the one for an agent that errored before emitting any transcript event, and isDeadProcess for a container that did not exit cleanly. A throw lands in the caller's catch, which records the error and writes no result file at all — so priorAttempts is a local that never escapes.

Scenario, on CI's default --runs 2: attempt 1 fails for a real reason and is recorded. Attempt 2's container is SIGKILLed, isDeadProcess fires, runOne throws, nothing is written. The attempt 1 transcript — exactly the kind of evidence #94 exists to keep — is gone.

Not hypothetical: the comments on those two guards cite 13 August and 21 August incidents where they fired across dozens of cells.

This is distinct from #93, which is about the upload step being skipped after a failing check.

Why it is not a one-liner

Writing the partial evidence to resultPath() would undo the invariant those throws exist to protect: an errored cell must not produce a result file, because the exporter reads that path and a wrong row published against a named vendor is worse than a missing one. That is the 13 August incident.

The exporter also globs <experiment>/*.json, so a sidecar next to the result file becomes a row with a mangled eval id. The evidence needs a location the exporter does not read — a separate top-level directory — and the workflow's artifact upload needs to include it.