fix: billed-cost reconciliation + no-pause speed keepalive

[jkbennitt]

Fixes the two remaining harness bugs from the v0.3.0 spread reconciliation (drafts 04 + 05 in results/spread/issues_to_file/).

1. Cost-estimate divergence (up to 4x vs OpenRouter actuals)

The token-count x price estimator missed in both directions on the spread (kimi 2.3x under, nemotron 4.1x under, grok 1.5x over). Root causes: per-provider reasoning-token response shapes, unmodeled caching discounts — and a genuine undercount where parse retries / failed deliberations never reached the tracker at all.

Fix — stop estimating, reconcile against ground truth:

• base_role._record_completion captures the generation ID of every provider call (incl. parse retries) into a per-agent buffer; drain_generation_ids() hands them off.
• The game loop drains all agents per tick (plus a final drain in run() for late timeout completions) into CostTracker.record_generation_id.
• fetch_billed_costs() queries OpenRouter GET /api/v1/generation?id= (the exact billed USD per call, semaphore-bounded, one retry for stats lag) and returns a frozen BilledCostReport. All-misses returns None, never a misleading $0; partial misses are flagged as a lower bound.
• run_scenario.py / run_benchmark.py reconcile automatically when the base URL is OpenRouter, write billed_cost into the summary JSON, and print it next to the estimate.
• analyze_spread.py passes it through as real_cost_usd/cost_source — same schema we hand-patched into the v0.3.0 leaderboard, so future spreads need no dashboard transcription.

The estimator stays for non-OpenRouter providers and live in-run tracking.

2. Pause-on-threat freezes slow models (--no-pause)

RimWorld force-pauses on threat letters; the loop only re-asserted speed at tick boundaries, so a slow model (kimi ~6 min/tick) left the game frozen for its entire deliberation window — frozen footage and wildly non-uniform game-time per tick across models.

Fix: a _speed_keepalive task re-asserts speed 3 every 10s for the whole run (no-pause mode only, speed_keepalive_s=0 disables, cancelled cleanly in a finally). Setting the speed is idempotent, and per-call failures are swallowed so a flaky RIMAPI moment can''t kill it.

Verification

• 513 tests pass (+22 new: billed-cost fetch via httpx.MockTransport incl. 404/connection-error/null-cost/auth-header paths, generation-ID capture/drain on the agent, keepalive loop/error-survival/cancellation)
• ruff clean, mypy src/ strict clean, --smoke-test green
• Live-game behavior (does speed=3 actually clear a threat force-pause?) to be validated on the next live run, per the usual caveat

🤖 Generated with Claude Code