docs: add 14-hour submission plan with time allocation and cut priorities

docs/SUBMISSION_PLAN_14H.md

@@ -0,0 +1,111 @@
+# Submission Plan — 14 Hours Remaining
+
+Date: 2026-03-07
+
+## Current state
+
+Done:
+
+- environment contract locked and stable
+- official constellaration verifier wired (low-fi run, high-fi submit)
+- 3 frozen reset seeds validated by measured sweep
+- reward V0 tested across 12 of 13 branches (replay playtest report)
+- random and heuristic baselines committed
+- PPO smoke passed on Northflank H100 (plumbing gate)
+- replay playtest script committed with full trace output
+
+Not done:
+
+- [ ] trained policy with demo-quality trajectories
+- [ ] HF Space deployment
+- [ ] Colab notebook
+- [ ] 1-minute demo video
+- [ ] README polish
+- [ ] paired high-fidelity fixture checks
+- [ ] submit-side manual playtest with successful high-fi outcome
+
+## Key finding: cross-fidelity gap
+
+The replay playtest (episode 5) confirmed that the canonical low-fi repair
+path from seed 0 crashes at high-fidelity evaluation. The state
+`(ar=3.6, elong=1.35, rt=1.6, tri=0.60)` is low-fi feasible but high-fi
+VMEC failure.
+
+Decision: **do not attempt to fix this in the remaining time**. Reasons:
+
+1. finding a high-fi-safe path requires sweep work with no time guarantee
+2. the plan doc already frames the trained policy as supporting evidence, not the product
+3. the gap itself is a strong honest finding for the submission narrative
+
+## Time allocation
+
+| Task | Hours | Notes |
+|------|-------|-------|
+| Train low-fi PPO | 2-3 | PPO smoke already passed. 24 discrete actions, 6-step budget. Target: visible score improvement across seeds. Run on Northflank H100. |
+| HF Space deployment | 2-3 | Hard requirement. Deploy FastAPI server, prove one clean remote episode. Debug dependency issues. |
+| Colab notebook | 1-2 | Connect to HF Space, run trained policy, show trajectory. Minimal but working. |
+| Demo video | 1 | Script around: environment clarity, human playability, trained agent, reward story. |
+| README and repo polish | 1 | Last step. Only after artifacts exist. |
+| Buffer | 2-3 | Deployment issues, training bugs, unexpected blockers. |
+| **Total** | **~11-14** | |
+
+## Execution order
+
+1. **Training** — start first because it can run while doing other work.
+   Use the existing `training/ppo_smoke.py` as the base. Train on all 3 seeds.
+   Stop when a few trajectories show clear repair-arc behavior (cross
+   feasibility, improve score). Do not overfit to one seed.
+
+2. **HF Space** — deploy while training runs or immediately after. The server
+   is already in `server/app.py`. Need to verify dependencies resolve on HF
+   infra and that one reset-step-submit cycle completes cleanly.
+
+3. **Colab notebook** — wire to the live HF Space endpoint. Load the trained
+   checkpoint. Run a short episode. Add minimal narrative connecting the
+   environment design to the trajectory evidence.
+
+4. **Demo video** — 1 minute. Structure:
+   - the problem (stellarator design is hard)
+   - the environment (narrow, human-playable, real verifier)
+   - human playtest (replay output showing legible reward)
+   - trained agent (trajectory with visible improvement)
+   - honest findings (cross-fidelity gap as a real insight)
+
+5. **README polish** — update with links to HF Space, Colab, and video.
+   Keep claims conservative. Reference the evidence docs.
+
+## What to cut if time runs short
+
+Priority order (cut from the bottom):
+
+1. Colab polish — minimal working notebook is enough
+2. Training length — a few readable improving trajectories over a long run
+3. README depth — link to docs, keep top-level short
+4. Do NOT cut HF Space — hard requirement
+5. Do NOT cut demo video — primary judge-facing artifact
+
+## Submission narrative
+
+Frame the cross-fidelity gap as a strength, not a failure:
+
+> The environment is instrumented well enough to reveal that low-fidelity
+> feasibility does not guarantee high-fidelity success. This is a real
+> challenge in multi-fidelity scientific design and exactly the kind of
+> insight a well-built environment should surface.
+
+The story is:
+
+1. we built a clear, narrow environment for one constrained design task
+2. we tested it thoroughly (sweep, baselines, replay playtest, 12/13 reward branches)
+3. we trained a policy that learns the low-fi repair arc
+4. we discovered and documented an honest cross-fidelity gap
+5. the environment is the product; the policy is evidence that it works
+
+## Risk assessment
+
+| Risk | Likelihood | Mitigation |
+|------|-----------|------------|
+| Training does not converge | Low — smoke already passed | Show the smoke trajectories as evidence. Document what was tried. |
+| HF Space dependency issues | Medium | Start deployment early. Have a local-only fallback with screen recording. |
+| High-fi submit never works | Already confirmed | Frame as documented finding. Do not promise high-fi results. |
+| Run out of time | Medium | Follow cut order above. Prioritize video and HF Space over polish. |