Wire #13 infrastructure into game loop + Docker e2e

[jkbennitt]

Summary

• EventLog emits at every tick stage (7 event types, thread-safe). CostTracker records token usage from agents after each deliberation.
• ActionResolver returns ResolverStats for conflict counting. MetricContext populated with conflicts + CentralPost message tracking.
• DEFAULT_WEIGHTS redistributed (proportional 20% cut) to activate coordination (0.12) and communication_efficiency (0.08). All 6 scenario YAMLs updated.
• --ablation runs 8 benchmark passes (full + 7 agent-removed).
• Optional Weave tracing on _call_provider().
• Docker e2e validated against real headless RimWorld — avg 0.804 across 6 scenarios, 99% parse rate.

Test plan

• mypy strict, ruff, 331 tests pass
• Smoke test produces events.jsonl + cost_snapshot + event_summary in benchmark_summary.json
• Docker: RIMAPI responds from host via socat bridge
• Docker: full 6-scenario benchmark (10 ticks, Nemotron 120B via OpenRouter)

Closes #13

🤖 Generated with Claude Code

[jkbennitt]

Code Review

Docker & Infrastructure

CRITICAL: Container runs as root
docker/Dockerfile:36-37 — USER rimworld is commented out. The entrypoint drops to rimworld for the game process via su, but socat and tail run as root. Privilege escalation surface if the container is compromised.

Socat bridge has no supervision
docker/entrypoint.sh:137 — socat is backgrounded with no PID tracking, no restart-on-crash, no error logging if bind fails, a race condition (healthcheck may hit before socat is ready), and no signal forwarding on container shutdown.

Healthcheck allows 30-minute wait
docker/Dockerfile:41-42 — 120s start-period + 30s × 60 retries = 1920s. Previous was ~2 min. 60 retries is overkill — reduce to ~20 retries (~10-15 min cap).

CRLF fix is good ✓
docker/Dockerfile:34 — sed -i 's/\r$//' prevents Windows line endings from breaking the script.

---

Orchestration & Scoring

MetricContext wiring — correct ✓
game_loop.py:413-431 tracks messages_acted_on, game_loop.py:450-451 accumulates ResolverStats for conflicts. Clean implementation.

EventLog thread safety — correct ✓
event_log.py:67 uses threading.Lock() with proper critical sections.

Weight redistribution — correct ✓
composite.py:10-21 — DEFAULT_WEIGHTS sums to ~1.0 within float tolerance. Tests validate via pytest.approx(1.0).

---

Agents & Scripts

Ablation logic — correct ✓
run_benchmark.py:386-387 iterates _ALL_AGENT_IDS, excluding one per pass. Clean one-at-a-time removal.

Weave tracing — truly optional ✓
base_role.py:163 initializes _weave_op = None, gated by None check. No hard imports.

Gap: ablation + save load failure
run_benchmark.py:399-402 — If load_game() fails and continues, agents run against stale state. Silent garbage ablation results.

Gap: --ablation lacks conflict validation
No check that --ablation and --runs / --no-baseline don't interact unexpectedly.

---

Tests

Good coverage on new features ✓

• TestResolverStats (5 test cases) covers zero-conflict through multi-conflict
• test_sum_to_one validates DEFAULT_WEIGHTS

Gap: scenario YAML weight sums not tested
test_scenario_loader.py checks structure but doesn't assert each scenario's weight overrides sum to 1.0.

---

Action Items

Issue	Severity
Docker root user	Fix
Socat supervision	Fix
Ablation error handling on load failure	Improve
Scenario YAML weight-sum test	Add
Healthcheck retries (60 → ~20)	Nit

Core instrumentation work is solid — thread safety, metric wiring, and weight math are all correct. Main concerns are Docker security and the ablation save-load gap.

[jkbennitt]

Fixes Applied

All review items addressed in feature/issue-13-wiring:

Docker security & reliability

• Root user: Removed commented-out USER rimworld. Entrypoint still starts as root for setup (chown, mv, ln), but socat and tail now run as rimworld via su
• Socat supervision: PID tracking + startup verification — exits immediately if bind fails. No more silent failures
• Signal forwarding: Added trap handler forwarding SIGTERM/SIGINT to all child processes (socat, game, Xvfb) for clean shutdown
• Healthcheck retries: 60 → 20 (~10 min cap instead of 30 min)

Ablation save-load guard

• Both ablation and main benchmark loops now continue (skip the run) when load_game() fails, instead of silently running against stale state
• Prints SKIP message so dropped runs are visible

Test coverage

• New TestScoringWeights — parametrized across all 6 YAML definitions, asserts scoring_weights sum to pytest.approx(1.0)

Verification

• 337 tests pass (6 new), ruff clean, mypy strict clean
• Need to re-validate Docker e2e before merge

[jkbennitt]

Docker e2e Validated

Rebuilt image with all fixes, ran crashlanded scenario (10 ticks, Nemotron 120B paid via OpenRouter):

Metric	Score
survival	1.000
threat_response	1.000
wealth	1.000
coordination	1.000
food_security	0.950
communication_efficiency	0.952
efficiency	0.889
self_sufficiency	0.667
mood	0.407
research	0.226
COMPOSITE	0.844

Stats: 95K tokens, 68 LLM calls, $0.03, 900s wall time

Entrypoint fixes confirmed working:

• socat bridge started with PID tracking (pid=146), no silent failures
• Signal trap registered for clean shutdown
• Save loaded successfully (game/load 200), all map endpoints 200
• Healthcheck passed with reduced retries (20 vs 60)
• Score climbed 0.824 → 0.844 over 10 ticks, 3 colonists alive

Ship it.