test(benchmarks): establish a 12-task main baseline

[minixalpha]

This establishes a baseline on twelve previously unrun development tasks using main at 2eb99b608dfb39464d17d109a4a8f11afda607a9. The frozen main agent passed 10/12 (83.3%); both zeros reached official correctness assertions. There were no final setup or verifier errors, no task replacements, and no whole-trial retries.

The reporting branch does not change agent or adapter behavior. It freezes task selection, source, prompt, budgets, and queue before model work, and preserves all installation failures alongside twelve accepted native install-only records.

Results

• db-wal-recovery: opening SQLite before backing up its WAL lost the original input. The agent guessed replacement values and checked those against its own reconstructed database; an actual WAL update was missing.
• pytorch-model-cli: C++ and Python checks shared an inferred 0.5/0.5 normalization. A post-freeze diagnostic confirms that plain pixel/255 matches all 50 official predictions, while the chosen normalization reproduces all 14 mismatches.
• The remaining ten tasks passed. PyTorch's verifier timed out once during dependency setup, then completed under the original 900-second per-attempt limit against the same agent output.
• Known token-estimated cost is $0.600974; one recovered model transport error lacks usage, so this is incomplete accounting, not an exact bill.

Full report · Machine-readable results and hashes · Registered configuration and preflight history

These twelve form a purposive development cohort alongside the earlier 22 tasks. The other 55 remain unopened by this experiment. One run per task does not establish reliability or improvement over another agent.

Infrastructure

• Add a reusable comparison orchestrator that uses the standard Harbor entry point and rejects changes to frozen plans or attempts to replay started trials.
• Add a verified original-image cache for slow public Docker Hub pulls: exact range validation, SHA256/size/DiffID checks, preserved OCI manifests, and RepoDigest verification after import. Task images, package sources, and official tests are unchanged.
• Allow reuse of successful native install-only trials only with complete task/source/wheel/bootstrap/config provenance. Both failed preflight jobs and all eight failed installation trials remain recorded. All twelve verifier dependency profiles are explicitly not_configured.

Validation

• 128 Harbor tests passed, including image-cache and accepted-preflight provenance coverage.
• All twelve official task suites ran; source snapshots, wheels, frozen trajectories, task files, configurations, pair barriers, and retry limits passed the final audit.
• Result counts, diagnostic mismatch counts, document links, and credential checks passed. Published artifacts contain derived observations and hashes only; raw traces remain local and require redaction before sharing.
• Harbor README changes are synchronized in English and Chinese. No bilingual research notes changed.

This remains a draft for review; no merge is requested by this update.