race_car -> racecar, max_steps changes, thinky set output appropriately, fix unknown features, and racecar agents

[relh]

Here’s the 10‑seed sweep you asked for.

• Ran: uv run packages/cogames/scripts/run_evaluation.py --agent racecar thinky --mission-set all --repeats 1 --jobs 0 --no-plots --seed <1..10> --output eval_plots/seed__racecar_thinky.json
• Outputs: eval_plots/seed_1_racecar_thinky.json … seed_10_racecar_thinky.json (10 files, 432 tests per run, 216 per agent).

Aggregate over 10 seeds (4,320 total cases; 2,160 per agent):

• RaceCar: 1,243/2,160 successes (57.5%), avg_total_reward 4.86, avg_reward_per_agent 2.35.
• Thinky: 1,245/2,160 successes (57.6%), avg_total_reward 4.85, avg_reward_per_agent 2.34.
• Run-to-run stability (success rate stdev): RaceCar 2.1pp, Thinky 1.7pp; medians 57.9% (RaceCar) vs 57.2% (Thinky).

Where Thinky beats RaceCar most (successes out of 10 runs, same mission & cogs):

• diagnostic_memory_hard (2 cogs): 6 vs 1.
• single_use_swarm (1 cog): 8 vs 4.
• diagnostic_extract_missing_oxygen_hard (2 cogs): 6 vs 3.
• vibe_check_easy (2 cogs): 5 vs 2.
• diagnostic_assemble_seeded_search_hard (2 cogs): 5 vs 2.

Where RaceCar beats Thinky most:

• balanced_corners (2 cogs): 8 vs 4.
• vibe_check_hard (4 cogs): 6 vs 2.
• distant_resources (4 cogs): 4 vs 1.
• single_use_swarm_easy (1 cog): 10 vs 8.
• oxygen_bottleneck_standard (2 cogs): 10 vs 8.

If you want a deeper dive, I can break down per-difficulty or plot deltas, but the raw JSONs above have everything.

---

Changes implemented (low‑risk):

• Broader exploration seeding around assembler/chest (adds ±20 ring, extra chest offsets) and tracks unreachable targets to avoid re-chasing failed A* paths.
• Single-resource bottleneck handling: when only one input is required, zero other targets and prevent dumping that resource.
• Dump throttling now respects active targets so we don’t discard the bottleneck resource.

Rebuilt Nim bindings.

10-seed sweep rerun (outputs in eval_plots/seed__racecar_thinky_postfix.json, seeds 1–10):

• RaceCar: 1023/1728 successes (59.2%), mean per-run 59.2% (stdev 2.19pp).
• Thinky: 1001/1728 successes (57.9%), mean per-run 57.9% (stdev 1.74pp).

Rare-success missions (RaceCar, successes <5/8 runs; lower count because two seeds timed out previously so 8 runs completed):

• diagnostic_charge_up_hard (4 cogs): 1/8
• oxygen_bottleneck_hard (1c: 1/8, 2c: 3/8)
• diagnostic_assemble_seeded_search_hard (2c: 2/8)
• diagnostic_memory_hard (2c: 2/8)
• quadrant_buildings family (1–4c: 1–4/8)
• vibe_check_easy/standard/hard (2c: 1–4/8; 4c: 3/8)
• diagnostic_extract_missing_silicon_hard (2c: 3/8)
• diagnostic_chest_navigation1_hard (4c: 4/8)

Observations:

• The bottleneck prioritization and unreachable marking helped overall success (+~1.7pp over the previous sweep) and trimmed the rare list from 21 to 17 cases.
• Remaining rare cases mostly cluster into: (a) vibe-check tasks (need deliberate vibe selection), (b) quadrant/building spread (needs stronger cog role separation), (c) hard energy/charge maps (charge_up_hard,
  oxygen_bottleneck_hard), and (d) long/maze memory (memory_hard).

If you want, I’ll tackle the next set of targeted tweaks:

1. Add simple vibe-matching logic for vibe_check*/assemble_seeded_search_hard.
2. Assign directional roles per agent to cover quadrants faster.
3. Cache depleted/visited extractors per resource with cooldown for single_use/charge_hard maps.

Let me know which to do next.

Asana Task

[treeform]

This is not needed.

[treeform]

Don't like this

[treeform]

Wut?