feat: unify conversational engine across interfaces

[ktjn]

Summary

• unify CLI, VS Code, and browser chat on one Python-owned typed conversation engine
• pass complete JSON schemas to Ollama and WebLLM so provider output remains valid structured plans instead of raw Modelable source
• replace the browser heuristic path with a semantic simulator and resumable conversation.* protocol
• make Suggest Projection use focused editor context, canonical rendering, exact multi-file previews, explicit apply/discard, and safe refinement failure recovery
• promote accepted compilation previews to the browser output/download collection without silently widening unsupported compile options

Root cause

The browser previously had a separate raw-source AI pipeline. WebLLM received only generic JSON mode, the generated union schema did not require every kind discriminator, and Suggest Projection bypassed the shared planner/context lifecycle. That allowed malformed or semantically incorrect output and caused interface behavior to diverge.

Validation

• uv run ruff format .
• uv run ruff check .
• mypy baseline ratchet: passed with 404 current errors and 45 resolved baseline entries
• uv run pytest --tb=short: 1547 passed, 21 skipped
• npm test (web): 356 passed
• npm run check (web)
• npm run build (web)
• npx playwright test: 40 passed
• npm run check:budgets (web)
• VS Code: check, build, 35 tests, and VSIX packaging passed
• opt-in Ollama conformance with qwen2.5-coder:14b: 6 passed
• strict MkDocs build passed

Doc/spec review: all phases passed