[kvlonge]

Why

M4i part 1: project resource assembly runs natively behind native-preview. Results match Python byte for byte, or native defers with a record.

Changes

• Rust assembly/project/: model, function and audit deps; managed source and seed destinations via the shared template evaluator; model and hook syntax checks for exactly the SQL Python validates.
• Python uses the facts, skips its own expansion and validation, and replays ENV/CTX reads into COMPILE_INPUT_READS. Anything else defers to unchanged Python with a project_assembly record.
• Review fixes:
  • The syntax check uses the shared parser-dialect guard (lineage/main/parser_dialect.rs). A dialect this parser build lacks (e.g. trino) defers with unsupported_dialect. New failure case assembly-hook-syntax-unbuilt-dialect.
  • ENV names are resolved by Python's os.environ.get and sent with the request, so Windows' case-insensitive keys match Python. Unsent names defer.
• Deleted: the assemble_native_project stub and assemble_project_by_engine seam. No oracle Python deleted.
• Changed expectations:
  • test_native_analysis_seams now records assemble_native_project_resources, which returns NativeProjectResources, not None.
  • _assemble_compiled_model drops seed_names, which model_build_deps never read.
• New tests: integration parity (Python calls 0, reads match), deferral cases and a simulated Windows environment; Rust tests checked against Python.

Verification

• Sabotage (reverted):
  • stage always defers → 6/6 fail;
  • managed source ignored → parity fails;
  • no read replay → parity fails;
  • no dialect guard → differential fails on the new case;
  • os.environ snapshot → Windows case fails.
• CI-flag differential: 321 identical. Preview deferrals: only syntax_error ×3 and unsupported_dialect ×1. validate wheel calls 176→22.
• Touched tests, compile E2E, differential E2E, make check-ci, make test, cargo test/clippy/fmt on 1.99.0, hygiene.