ui: chatbot dark-neon theme — readable bubbles on Watch Agent Play

fix(ui): drop unsupported type= kwarg; dict format works by default

fix(ui): Gradio chatbot format error on Watch Agent Play

UI: fix '0 runs' chip + instant tab-card flip via js= callback

UI: replace tab strip with big-card navigation

Semantic run names: Probe/Drift/Anchor/Restrain/Champion + regen all plots

Reframe: drop Run 7 from hero, keep only where appropriate

Polish Blog.md: fix broken images + visual hierarchy

Stronger opening: meeting scheduling with 3 fabricated fields

Hero Zone redesign + fix Run 7 missing from plot 08

Expand Blog.md to comprehensive deep-dive (5.9k words, all evidence)

Reframe: environment is the contribution, training is validation

Add full plot deck to Blog and README

Submission polish: Run 7 headline across all surfaces

Run 7 BEATS BASE (+19%) + UI fixes

Blog: human hook opening, simple language

Judge-ready polish: diagram, before/after, curated replays

Fix neon CSS injection for Gradio 6.x + Blog.md at root

Update blog + slides with Run 5-7 findings

Port 7860 + neon theme

Neon cyberpunk theme - dark bg, glowing accents, stat cards

Fix Gradio compatibility for openenv's bundled version

Add rich 4-tab Gradio UI dashboard

Run 6 results + training fixes + all plots regenerated

Align eval prompts with training: add required_keys to initial context

chore: track remaining PNGs via LFS/xet

plots: add training progression + diagnostics, drop W&B links

docs: README aligned with hackathon judging criteria (Judges 60s tour + storytelling arc + plot captions)

docs: sync README.md (slide deck + auto-validator gate update)

docs: sync SUBMISSION_CHECKLIST.md (slide deck + auto-validator gate update)

docs: sync docs/slides.md (slide deck + auto-validator gate update)

docs: add detailed model cards for Run 1 / Run 2 / Run 4

docs: sync docs/trace_demo_run1.md

docs: sync docs/trace_demo.md

docs: sync docs/STATUS.md

docs: sync docs/blog.md

Add plots/ for inline embedding in env Space README

Sync canonical README (KL-anchor narrative, embedded plots, deliverable links)

env: bump max_concurrent_envs from 8 to 64

eval: enforce one-tool-call response format on every turn

Fix parser: handle quoted commas, balanced parens, ASK:/PROPOSE: prefixes

Eval system prompt: align character-for-character with training PROMPT — ensures trained model has zero distribution shift between train and eval

Eval system prompt: drop misleading software-stack example, align with training PROMPT (forces model to use task-family fields, not copy the example verbatim)

Parser: support ASK:/PROPOSE:/Q:/PLAN: prefix forms produced by Qwen3 GRPO

inference: parser fix — handle key=value in func calls + balanced parens

fix(eval): pass enable_thinking=False to disable Qwen3 thinking + bump MAX_TOKENS to 800

feat: add run_eval.py to Space (needed by eval_with_vllm.py for trained-model evals)

env: enable concurrent rollout sessions (max_concurrent_envs=8)

rewrite training notebook with cleaner cell-by-cell structure

Add training/train_grpo.ipynb — GRPO training notebook (TRL + vLLM + ClarifyEnv)