[kvlonge]

Why

Preview still called the wheel for input enrichment and pivot proofs; both move native.

Changes

• Enrichment runs natively via analyze_query with Python's projection and lineage. It defers wherever Python leaves its fast path: CTE fact recovery, NULL filters, normalisation or analysis failure, non-select shapes, casts with no upstream column, non-ASCII function names.
• Pivot proofs run natively, in the session and via the seam for unanalysed models. Parse errors and non-ASCII names defer (dynamic_pivot).
• Rust pivot cases; three failure-corpus cases. Wheel-site report test (rebased on feat(native): run semantic checks natively behind the preview engine #1050): preview expects only metadata_checks deferrals, as input_enrichment no longer occurs.

Test expectations added (parity test):

• Two contract-enforced CTE models where Python's CTE recovery changes types; without the deferral check they mismatch.
• input_enrichment deferrals <= 120 (now 104): every-enrichment-defers fails.
• Native enrichments >= 90 (now 106), counted at native lineage, so fallbacks do not count.
• Pivot proof minimums for the new generated shapes.
• New: a pivot outside --select analysis is proven natively, with no wheel call.

Sabotage, run and reverted: deleting the CTE check gives 3 mismatches; all native enrichments deferring gives 495 deferrals; a monkeypatched native session gives 0 native enrichments; a monkeypatched native pivot proof hits the wheel.

Verification

• cargo test --workspace --all-features, clippy, fmt (1.99.0) before the rebase.
• Before the rebase, make compiler-differential (CI flags): 302 identical, enrichment and dynamic_pivot wheel sites 0, 12 legacy_analysis deferrals. -shipped: 293 identical. tests/e2e/scripts/compiler_differential passes.
• make check-ci, make test, hygiene.

Residuals:

• No corpus project hits the CTE-recovery deferral; generated parity projects do.
• Pivots fall back to the wheel when the whole session hands back.
• Any non-ASCII name defers every pivot proof.