Run the 200-episode StatefulPuzzle grid + Phase 1 report

[pixiiidust]

What to build

Execute the full Phase 1 grid and produce the SC/DD surface report.

Run 200 episodes (SC 5/10/15/20/25 × DD 1/2/4/6, T=40, 10 episodes/cell) on deepseek/deepseek-v4-flash with both caps live (est. ~24,000 calls, $10–20, cap $50). Then produce the Phase 1 report: all four heatmaps, collapse-onset table, token/cost summary, and at least five sample traces where world state diverges before the first invalid action. Evaluate the plan §9 acceptance checklist explicitly (visible gradient/cliff, ≥10 paired collapse episodes, pct_ws_first > 0.50, no single API/JSON failure mode explaining the result). If heatmaps come back flat, follow the plan §10 playbook and file the findings on this issue instead of the report.

Reference: docs/plan.md §6, §7, §9, §10.

Acceptance criteria

• 200 episodes completed, or a clear cost/API blocker documented on this issue
• Total spend within the $50 cap; token/cost summary included
• Phase 1 report with all §7 outputs committed under the configured output paths
• §9 acceptance checklist evaluated line-by-line in the report
• Parse-failure audit: json_parse_failure does not dominate the collapse region

Blocked by

• Config-driven grid runner (U3) #4
• Grid analyzer with ESL analysis, label + realized (U4) #5

[pixiiidust]

Boundary probe verdict (pre-grid): 4 episodes at SC=40 (DD 1/6, T=40) cost $0.24 and found the transition band: success 1/4, mean WSA 0.86–0.93 vs ~0.997 at SC=25, action validity down to 0.63. No truncation artifacts (0 ceiling hits at 16k; fallback ≤8%). Grid extended to SC [5,10,15,20,25,40] → 240 episodes, launched with --updater-max-tokens 16000 (probe peak updater output was 9.3k, too close to the 10k default).

Analysis note for the report: at SC=40 no step crossed WSA < 0.5 (min 0.81) while invalid actions appeared as early as step 2 — aggregate WSA stays high while enough facts are wrong to break preconditions. The Phase 1 report should include a wsa_threshold sensitivity pass (e.g. 0.9) alongside the plan §7 definition, per plan §10 Case D guidance.