Merge pull request #5 from yiyu-tian/scenarios/tech-production-planning

botgen: surface scripted opponents for YAML; drop premature capture_actor

mapgen: surface map generator for YAML (base_map generator spec)

win-speed bonus: record + reward how fast a model wins

tests: tech-production-planning pack — full Rust loop

#2 hard: set base_map INSIDE overrides so it actually uses scout-arena

#2: enforce ordered waypoints (stateful waypoint_sequence) + two-path medium + scouted hard

action-multiunit-coordination hard: spatial-grounding via relative-direction objective

prompt_v2: sanitize confusing 'Buildings: ? ()' briefing line

action-multiunit-coordination medium: disperse the 3 regions (multi-vector)

Default spawn_mcvs:false — stop engine auto-seeding phantom MCVs

playback: append a terminal 'episode end' frame (the resolved win/loss board)

Fix action-multiunit-coordination: enforce split, binding deadline, real fail

De-dup: consolidate aimless scenario duplicates (24→21 active)

Scenario configs (level+fog per cell) + adversarial-duel duel/interrupts + clearer names

Structured-fog text mode, premium routing, codex descriptions, minimap colour-by-difficulty

prompt_v2: strip ASCII minimap from the briefing (training parity)

test: pin run_id in whole-result determinism test (wall-clock label, not part of contract)

Wire bench to vendored training prompt v2 (system/briefing/minimap)

Vendor training prompt/briefing/minimap v2 (byte-identical)

Training-parity minimap (real terrain + legend) + viewer (system/thinking/debrief)

Live-smoke fixes: tool-call wire 400, episode resilience, real PNG minimap

run_eval: dependency-free .env autoload for the API key

Deterministic scenario-scoped game knowledge + explicit objective

Strategy packs: faithful 'destroy key economic buildings' objective

S8: capture_actor tool schema + dispatch; wildcard count 21->22

Hard-tier curation: action/perception/reasoning/strict/rush-hour + exhaustive classification

Hard-tier curation: Art-of-War quartet

Hard-tier curation: adversarial family (skirmish, siege)

Hard-tier curation: seed-driven spawn variation (exemplar + invariant)

Curate: remove the auto-generated cat-* corpus

Eval resilience layer for real OpenRouter runs

Scenario hygiene: quarantine redundant cat-* + harvest packs

Art-of-War long-horizon family: sacrifice/detour, delayed credit

Strict-action-API family: spec fidelity as the objective

Adversarial 1v1 spotlight: ladder family + rating + Elo wiring

cargo tools: enter_transport + unload, 1:1 congruence

Unified Battle Viewer in app.py + run/model playback identity

guard tool: 1:1 congruence with engine Command.guard

Wire goal tracker into scoring + leaderboard

Playback: capture model reasoning + per-turn goal tracker + viewer

Scenario-controlled tool allow/deny + default core set

S7 bench: surrender tool + loss outcome (tool schema 1:1, 15==15)

Step 4 (bench): interrupt-driven loop + playback/trace capture

Step 2: concurrent scenario execution (--concurrency N)

Pipeline step 7: per-episode playback persistence

Bench: surface S9 spatial tensor in render_state (multimodal reach)

Pairwise adversarial eval + Elo (the user's 'pairwise conditions')

Generalization-gap metric: held-out split in run_eval + leaderboard

Catalog C13: harvest-economy packs — closes #14 (user economy families)