Surface RIMAPI action errors to agents + fix load-settle race

[jkbennitt]

Summary

Two fixes from live-game testing (2026-05-16 smoke run on Nemotron 120B-A12B, Crashlanded) where agents were re-proposing the same invalid action every tick because RIMAPI's 500-on-semantic-error was opaque to the agent loop.

• Load-settle race in run_scenario.py: polling broke on first population > 0 (~4s after load_game), but RIMAPI returns 200 before Unity finishes applying the load. The immediately-following unforbid_all_items() POST then got 500'd. Now requires 5 consecutive stable-population checks (10s floor) before writes.
• Action error feedback loop: ExecutionResult.outcomes now carries an ActionOutcome per dispatched action with success + cleaned-up error message (unwrapped from RIMAPI's {"errors":[...]} JSON envelope). game_loop broadcasts STATUS_UPDATE with failed outcomes via CentralPost, so agents see "DO NOT REPEAT" context next tick. Removes the "researcher keeps re-proposing already-finished Electricity" loop observed live.
• Pre-existing positional telemetry bug fixed: i < executed had assumed actions failed in order; now uses per-action outcome flags.

Diff: 4 files, +167 / -13. 382 tests pass; ruff + mypy strict clean.

Why this matters for the benchmark

Agents currently can't learn from their own mistakes within a run — they see opaque server errors, retry the same invalid combo, and burn deliberation budget on no-ops. The ActionOutcome + CentralPost broadcast is observability and an agent learning loop in one piece. It's the first piece of the observability floor in the broader Phase A restructuring.

Test plan

• pytest — 382 pass (4 new outcome-capture tests)
• ruff check src/ tests/ scripts/ — clean
• mypy src/ — strict clean
• Re-run the same live smoke config (Nemotron 120B-A12B, Crashlanded, 10 ticks) on this branch to verify mood (0.408) and research (0.226) move under the new feedback loop — tracked as A10 in the project plan.

🤖 Generated with Claude Code