Build an information-sensitive live SC2 belief benchmark

[jwm-dev]

What changed

• Replaced the weak generic macro-intent test with a full-game TvZ hidden-commitment benchmark.
• Added one shared-opening Zerg opponent with roach/ravager, mutalisk, and ling-flood branches that share all branch tech.
• Added irreversible anti-armor, anti-air, and anti-swarm doctrines plus an auditable legal scanner-sweep action.
• Added deterministic current-observation and privileged-state experts and a recurrent legal-history controller trained from privileged action advantages.
• Extended full-game evaluation to 1,500 game seconds, added paired tournament/summarization tooling, committed the frozen checkpoint and aggregate artifacts, and added no-leak regression tests.
• Rewrote REPORT.md around the controlled result and its limitations.

Result

Across six paired held-out contexts:

• observation: 3-0-3, score 0.500
• recurrent belief: 6-0-0, score 1.000
• privileged oracle: 6-0-0, score 1.000
• belief without scouting: 3-1-2, score 0.583

The recurrent controller recovers 100% of the observed oracle gap on this small mechanism test. The clean separation is concentrated in the mutalisk branch, so the report recommends a larger uniformly action-sensitive replication before architecture tuning.

Validation

• .venv\Scripts\python -m pytest -q: 14 passed, 1 skipped
• .venv\Scripts\ruff check .: passed
• Frozen model loads in the bot environment (64 hidden units, 100 legal features)
• Aggregate summary and figure regenerated from the four raw phase directories
• git diff --check: passed