Reintegrate as a coral-compatible backend for the DealiiX platform

[pcolt]

Closes #12

Overview

Reintegrates coral-python as a coral-compatible backend, so the DealiiX platform can drive it by only
changing the executable/plugin paths (it stays backendKind: coral on the platform side). This delivers the
coral-python half of the local MVP; the platform-side changes are in 2listic/dealiiX-platform#209.

phiflow-obstacle-fe-coral-py-low.mp4

Summary

• chore: adopt a uv project layout (pyproject.toml + uv.lock); docs updated to uv sync / uv run / uv add.
• feat: coral-compatible CLI in main.py — a global -p/--plugin option (comma-separated modules; empty = all)
  plus register / run subcommands, matching the C++ coral surface. A self-locating coral-py launcher runs it
  inside the uv project while preserving the caller's working directory.
• feat: generate_registry now emits the platform-native registry — keyed by node type, with primitives
  carrying an empty arguments list. method_name and is_valid are dropped (parity with C++ coral, which emits
  neither — the editor sets is_valid itself and identifies nodes by type).
• feat: the executor treats type as the sole node identifier and infers the kind
  (primitive/function/constructor/method) from the definition maps (_classify), so it runs the platform's lean
  graphs (no node_type/method_name per node).
• test: all workflow fixtures normalized to the lean shape; registry/executor tests updated; removed the unused,
  outdated invalid_workflows fixtures. 88 tests pass.
• docs: add docs/ONBOARDING.md (+ Italian docs/ONBOARDING.it.md) — an onboarding/architecture guide for
  colleagues, written for two personas (add a CFD/scientific library / improve internals & the platform
  contract). Includes an honest assessment of why inspect.signature is used (with the alternatives), and how
  the executor orders execution and classifies lean nodes (_classify). Referenced from README.md / CLAUDE.md.
• fix: restore the standalone phi_flow/one_obstacle.py example (dropped in the examples reorg) and correct its
  run path in README.md; refresh the stale module-registration docs (_MODULES) in definitions/README.md
  and CLAUDE.md.
• docs: document ffmpeg as a prerequisite (README.md Installation section + phi_flow/README.md) — required on PATH for .mp4
  export from the PhiFlow scripts and the phiflow_plot_and_save workflow node (matplotlib's
  anim.save(..., writer='ffmpeg')).

Verified

• coral-py -p "math,string,phiflow" register writes a node_types.json the platform accepts (Save & Sync
  populates the sidebar with no skipped nodes).
• coral-py -p math run <lean graph> executes a platform-style lean graph end-to-end.
• In-app: built graphs and, from the platform, loaded and ran every graph in
  tests/fixtures/valid_workflows/ (including the phiflow ones) via Execute (local) — they run through
  coral-python and produce output.

Test plan

• uv sync then uv run pytest — all green.
• uv run python main.py -p "math" register produces a type-keyed node_types.json.
• uv run python main.py -p "math" run tests/fixtures/valid_workflows/network-from-fe-math.json runs and prints results.

Deferred (future work, not in this PR)

Remote/Slurm execution, pipeline stages, per-node --touch-dir status, and promoting coral-python to a git
submodule when containerizing. Track these separately once #12 is closed.