[kvlonge]

Why

Under native-preview, native model contract evaluation took ~15.7 s CPU on the dense 3,000-model project, against ~0.2 s for the Python engine (the same on main). That made the whole preview compile ~2.4× slower and blocks promoting the lane.

Changes

• type_system/_helpers/parsed.rs: is_python_expression returns false for scalars before trying to deserialize them as Expression. Every Expression variant carries data, so a scalar can never deserialize. The failing deserialization built a large error message each time, costing ~100 µs per normalize_type call; now ~5 µs.
• contracts: a request-scoped TypeComparer normalizes each type string once, as Python's lru_cache on normalize_type does. No global state.
• compile_timings gains contracts_cpu_ms (CPU time of the calling thread).
• Same-runner ratio tool: --base-engine/--head-engine and --gate-phase (gate one reported phase, plus --phase-noise-floor-ms). A new matrix entry in the ratio job gates native-preview contracts_cpu_ms at ≤1.00× python + 50 ms on dense 3,000 models.
• Ratio tool: a phase missing from any run is reported missing, not counted as 0.
• Changed expectation: test_compile_json_behavior timing keys include contracts_cpu_ms.

Verification

• Dense 3,000 cold, 3 alternating runs: contracts CPU 124 ms python vs 90 ms native-preview (was ~15.7 s).
• Guard sabotage (reverted):
  • original slow path: 14,771 ms, fails;
  • memo removed only: 334 ms, fails.
• New Rust test: a batch sharing types equals per-model requests, including repeated deferrals.
• Contract and type-normalization parity tests, ratio-tool E2E, test_compile_json_behavior.
• CI-flag differential: 320 identical, no contract deferrals.
• cargo test/clippy/fmt on 1.99.0, make check-ci, make test (11938 passed), hygiene.
• No local real project compiles against 0.151.0, so dense parity and timing stand in.

[kvlonge]

Landed together in #1075 to cut merge-queue CI runs; review history stays here.