fix: reuse cached results for single-experiment runs

[benjamincanac]

Summary

agent-eval <name> (run one experiment) routed to a separate code path from agent-eval / run-all, and the two disagreed on how results were stored and reused. The single path always nested results by model, never wrote a fingerprint into summary.json, and never scanned for reuse. Net effect: results produced by the documented "run one experiment" command were structurally invisible to resume, so they got re-run on every invocation and the next run-all re-ran everything from scratch.

• Default action now delegates to runAllCommand, so the single-name case is just run-all filtered to one name. It computes fingerprints, scans for reuse, passes fingerprints into runExperiment, and stores single-model results at results/<name>/<ts>/.
• Removed the now-dead runExperimentCommand and helpers it owned (resolveConfigPath, plus unused imports).
• Hardened scanReusableResults to recurse through model-nested directories, so legacy results at results/<name>/<model>/<ts>/ (and normal multi-model layouts) are reused without a manual backfill. The fingerprint encodes the model, so cross-model false reuse can't happen.

Repro (before)

agent-eval my-experiment            # runs, saves results
agent-eval run-all my-experiment --dry
# => "25 to run, 0 cached" even though everything just passed

After the fix, the second command reports the just-passed evals as cached.

Behavior notes

• Single-experiment accepts bare names, globs, and path-style experiments/cc.ts (normalized to the bare experiment name), so existing invocations keep working.
• Fixtures are loaded per experiment, so the validation config field (e.g. validation: 'none') is honored by the unified path — previously only the single-experiment path respected it.
• An explicitly named experiment that can't run (config fails to load, missing API key, no valid fixtures) exits non-zero instead of being skipped with exit 0. No-args sweeps still skip unloadable configs, but an experiment with no valid fixtures fails the sweep's exit code.
• Infra/timeout failures are still re-run (isNonModelFailure behavior unchanged).

Test plan

• agent-eval <name> then agent-eval run-all <name> --dry reports cached (new cli.test.ts case)
• scanReusableResults picks up single-segment and provider/model nested layouts, and prefers newest across mixed layouts (new results.test.ts cases)
• Path-style experiments/cc.ts argument still works (new cli.test.ts case)
• validation: 'none' evals without EVAL.ts are discovered (new cli.test.ts case)
• Named experiment with invalid config or missing API key exits 1 (new cli.test.ts cases)
• npm run lint, npm run build, npm test (221 passed, 14 skipped)
• changeset added (patch)

[benjamincanac]

Closing as code as diverged and it can be fixed by using run-all.