Submission polish: refresh site to v0.3, drop internal working docs, clean output tone

[duckyquang]

Summary

Went through the whole repo to get it ready for the public + the AAAI code drop. Mostly cleanup: refreshed the landing page to v0.3 with the real repair-ranker numbers, pulled a bunch of internal working docs that shouldn't be public, and cleaned up tone/leftovers in the code and docs. Net is +90/-909.

The bigger pieces:

• Website was stuck on v0.2 — bumped to v0.3, swapped the "clean numbers pending" placeholder for the actual results (0.997 nonlinear, 0.380 held-out linear, 0.420 transfer), fixed the entrapment CI (was ±0.086, should be ±0.109) and a dead OUTLINE.md link.
• Removed internal-only docs: FIXING_PLAN.md, STYLE_KIT_STATUS.md, the stale overnight SUMMARY.md (it also carried a /Users/sparsh/... path), and REVIEW_ATTACKS.md (see notes).
• Bannered the superseded result docs (roadmap / p1_baselines / h2_report / learned_solver_fix) so the old value-diffusion numbers don't read as current.
• De-internalized TECHNICAL_GUIDE + RUNBOOK (dropped "read before P0", per-person seed assignments, "ping the team") and pulled the emoji + "FLAG THE TEAM" strings out of the noise ablation + plot output.

Notes

@ImSpxrsh @dyin08 @Gun-Akash a few things I'd like eyes on:

• I deleted paper/notes/REVIEW_ATTACKS.md — it's our rebuttal prep (the "paper is borderline" read, reviewer personas, defusals). It was public and not excluded from the supplementary build, so it would've shipped to reviewers. It's still in git history if we want it privately — say the word if you'd rather I stash it somewhere than delete it.
• Left the display font (Fraunces) as-is. A design linter flagged it as generic but I think it's fine and didn't want to churn the look — easy to change if you disagree.
• Kept the "a large language model assisted with readability" line in the paper — AAAI requires disclosing that, so it stays.
• Open question I hit while cleaning: the AAAI-27 page limit isn't confirmed, and the AI-use statement spills ~4 lines onto an 8th page under the proxy build. Worth checking before we submit.
• Folded in the CHECKLIST wall-time change that was already sitting uncommitted (the dimension-scaling ~2.3h measurement).
• All 431 tests still pass locally.