another-harness Run-History Decision

another-harness has now resolved the first honest storage fork created by its live builder/evaluator loops.

Main architectural facts

The repo explicitly decided not to make state/runs/ canonical yet.
Canonical operational truth remains in:
- state/work-items/*.json
- state/evaluations/*.json
- state/handoffs/*.md
Historically important attempts may be preserved as derived evidence bundles under docs/plans/artifacts/<work-id>/<attempt-id>/....
The preserved 0028/0030 live evidence is already mechanically replayable through tools/check_real_loop_smoke.py, so there is no need to invent a second truth plane merely to rescue those attempts from transcript oblivion.
The new tools/check_run_evidence.py checker validates the storage decision itself, the preserved bundle structure, the actual replay inputs (builder.patch and evaluator-artifact snapshots), and the fact that replay still works without preserved run.json / episode.json metadata.

The repo now has a thinner answer to run history:

keep canonical truth in the work/evaluation/handoff loop
promote only historically important attempts as evidence bundles
reopen state/runs/ only if future many-attempt pressure or resume/recover work genuinely earns it

This is a design gain precisely because it refuses bureaucracy before necessity. A rare civic virtue.