refactor: replace v3 workflow with ChatGPT + Codex thin core

[Moritz0924]

What changed

• archive the v3.2 workflow at v3.2-pre-thin-core
• add the eight-stage v4 state machine, strict five-gate transitions, immutable file handoffs, context hashing, migration support, and unified CLI
• add independent ChatGPT and Codex result-contract prompts under prompts/formal_v4/
• freeze and hash-protect the existing 16-stage prompts without modifying them
• replace the dashboard with a lightweight formal-workflow console backed only by the CLI
• migrate the current formal competition assets to v4
• remove the training sandbox, local model runners, scoring/training policies, vendored training skills, and legacy execution chain
• update architecture, onboarding, CI, and operational documentation

Why

The previous mainline mixed the formal competition workflow with training, local-model routing, and learning-platform features. Model quality and compute varied across providers, which increased state divergence and made evidence harder to reproduce. The new design keeps ChatGPT as the modeling and writing lead while Codex owns local data, implementation, reproduction, and numerical verification.

Impact

• one competition maps to one workspace
• no model API is required by the formal workflow
• ChatGPT and Codex exchange immutable, hash-bound files
• only the controller advances state; only humans confirm gates
• implementation conflicts can be fixed locally, while model or conclusion conflicts return to ChatGPT
• the v3.2 system remains recoverable through the archive tag

Validation

• python -m unittest discover -s tests -p "test_*.py" -v: 26 passed
• python -m compileall mmwf scripts tests: passed
• python scripts/validate_contracts.py --stage current: passed
• python scripts/check_gates.py: passed
• npm ci and npm run build: passed
• full simulated eight-stage workflow, including five human gates: passed
• legacy prompt hash protection: passed
• high-confidence secret scan on the new mainline and archive: no findings

Local pytest was unavailable in the current Python environment; CI installs requirements.txt and runs pytest -q.