run_eval: add --bedrock-region flag and route bedrock provider config

Bucket D: discover archived packs + load adversarial-siege from _archive

Phase 4 fix: run_eval double-nest in full_playback dir

Quality drive: schema fix, 5 new/revised packs, 4 engine tests, scenario audit

Paper-experiment prep: human-study harness + pass^k + paper plan

Add handoff ablation (recover-from-deficit / capitalize-on-advantage)

Add perception ablation grid (observation channel × fog of war)

feat(providers): add together.ai preset (Qwen3.6-Plus and friends)

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

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

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

run_eval: --or-provider parses provider/quant correctly

run_eval: --or-provider (pin OpenRouter provider/quant, no fallback) + --fog-mode CLI

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

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

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

Eval resilience layer for real OpenRouter runs

Scenario hygiene: quarantine redundant cat-* + harvest packs

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

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

Wire goal tracker into scoring + leaderboard

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

Pipeline step 7: per-episode playback persistence

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

#6 leaderboard: data layer + run_eval publish + Gradio tab

Add scoring + P/R/A diagnostics + run_eval CLI