[kvlonge]

Why

Under native-preview, semantic metadata checks still ran in Python, and B102/B301 UDF checks parsed models with the Polyglot wheel. This finishes the native semantic-checks stage.

Changes

• Native metadata checks: B102 arity and B301 argument families for UDF calls; B300 config references (incl. cursor_inputs) and B301 cursor types; B300 source cursor_column; B302 SQL-test columns; Python's positions
• Python passes data in and assembles diagnostics in Python's order. Non-ASCII folding, unsupported dialects/types or native failures fall back to Python's checks and record a deferral
• Kept in Python as stage inputs (for M4l/M5): expression_is_true, relationships, accepted_values audits and resource SQL checks. No Python is deleted: the python engine and the fallback still use the metadata path
• _helpers/ regrouped into four folders (pure move)
• E2E file renamed test_native_semantic_checks_cli.py: its module name clashed with the integration file, dropping tests in combined runs

Changed expectations:

• Harness wheel-site cases, the semantic CLI E2E and the deferred-semantic integration case drop metadata_checks metadata_validation.py, which no longer occurs. input_enrichment analysis_session stays
• New: a parity test, a CLI E2E case, 3 failure-corpus cases, and a unit test keeping SEMANTIC_CHECKS no higher than TYPE_SYSTEM

Verification

• Sabotage: forcing native metadata checks to defer fails the parity test, both CLI E2E cases and the harness test; reverted
• Rust tests; parity on generated projects in 4 dialects for all B102/B300/B301/B302 families, none deferred
• make compiler-differential (CI flags): 307 identical, coverage 56/56, no semantic deferrals, semantic wheel sites 0
• Compile E2E for seams, lineage and semantic checks, plus tests/e2e/scripts/compiler_differential: 300 passed
• cargo test --workspace, clippy/fmt 1.99.0, make check-ci, make test, hygiene