collect_eval_data: cell-level retry on transient API errors

Phase 5: retract F1 (Plus passivity) — Together adapter bug

Rewrite scenario descriptions in natural mission-brief prose

Phase 5 update: 112-cell snapshot, gemma 56% wins, Plus 0/55

Phase 5 doc-coherence pass: sync all per-cell counts to 93

Phase 5 closeout: 93-cell final table, gemma 6/6 + Plus 39 cells

FINAL_STATUS: session checkpoint with final 86-cell per-model numbers

data/maps/: refresh oramap binaries from session runs

data/runs/: commit run metadata; gitignore bulky JSONL/PNG payloads

Phase 5 final 83-cell table: gemma-31B (40%) vs Qwen3.5-9B (42%) vs Plus (0%)

Phase 5 nuance: gemma-31B WINS spec-tanya-c4 → Plus passivity is MODEL-SPECIFIC

Phase 5 final: 74-cell triage strengthens F1 to ≥30pp scaling delta

FINAL_STATUS: session continuation — defensive cash fix + gemma run

Defensive fix: strip agent.cash=0 when starting_cash>0 in tmp YAML

FINAL_STATUS.md: comprehensive summary of all 5 phases

Phase 5 final: PHASE5_FINDINGS.md, headline F1 = scale-INVERSE passivity

Verbosity sweep complete: 97.1% of descriptions trimmed to <=40 words

Fix mfb intended-policy tests broken by per-player cash regression

Map diversity: 11 packs rebound to custom maps (was rush-hour-arena)

Phase 5: F1 deeper triage — Reasoning not Action

Phase 5 rolling findings: F1/F2/F3 from 19 paper-v1 cells

Audit cleanup: archive 3 packs flagged in SCENARIO_UNIQUENESS_AUDIT

Fix P0 regression: explicit agent.cash in 3 test packs

Phase 4 + Phase 5: smoke verified, real collection running, F1 finding draft

Phase 1 engine audit: ENGINE_AUDIT.md + bench-side closures

Phase 4 fix: run_eval double-nest in full_playback dir

Phase 4 data plumbing + 3rd defect partial fix

Defect fix: 2 of 3 packs flagged by uniqueness audit

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

Engine-feature integration: 4 commands + 9 scenario packs + 4 test suites

Add triage tool + scope paper around out-of-engine features

Defect-fix wave: restore the no-cheat bar across 10 packs

Add scenario quality audit + coverage map scripts

Strip empty tool_calls from wire history (Together Qwen3.6-Plus fix)

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)

Persist human Play-tab runs in the standard Playback format

Vendor the necessary openra_rl_training + openra_env modules

Bundle canonical map assets in-repo (addresses PR #12 map concern)

Phase 2: Play tab — minimap fills its space (drop fixed height)

Phase 2: Play tab — contextual clicking + per-type unit shapes

test: HumanController drives a 1v1 side (Phase 2 x Phase 3)

Phase 2: Play tab — selection boundary, move arrows, clear selection

Phase 2: Play tab — exact-cell unit selection with toggle

Recalibrate 8 packs regressed by the count: spawn-spread engine fix

Phase 2: Play tab — fix hp, reorder panel, cleaner units table

Phase 2: fix Start button + tactical minimap (shapes/legend/counts)

Phase 2: Play tab — larger minimap + bigger coordinate labels

Phase 2: Play tab — fix objective strikethrough + add coordinate grid