final changes

serve /demo dashboard + episodes as static via FastAPI

reward shaping v4: break the no-commit attractor

phase-2 escalation: worked-example prompt + step cap 12->8

phase-2 tuning: anti-no-commit prompt + REINFORCE baseline env var

phase-f fix: phase-1 domain coverage + checker robustness + flush + runs persistence

phase-e fix: align eval pool ground truth with checkers (holdout templates solvable)

phase-e fix: random baseline answers with last tool_return + eval pool coverage

phase-d: 60 task templates x 6 domains + programmatic checkers + full loader

phase-c: wire env action loop + real /reset /step /state in app.py

phase-c: 5 rubrics + weighted total + hard-compromise floor + calibration ladder

phase-c: injection templates + compliance predicates + interim corpus loader stub

phase-b: author 50+ content tools across 6 domains + leak sinks + eval holdout

phase-b: catalog generator with rotation + bait + requires_tools enforcement

phase-b: paraphrase sampler + offline paraphrase generation script

phase-b: name namespaces + flag-independent uniform sampler

phase-b: content pool registry + phase-keyed shape invariants

phase-a: env + client stubs (client respects no-server-import rule)

phase-a: fastapi stub with validator-formatted stdout markers

phase-a: repo skeleton + validator-aligned manifests