[kvlonge]

Why

Under native-preview, model analysis took 13.0-13.5s CPU on dense 3,000 models vs ~4.2s for Python. The session analysed every model alone; Python shares one analysis across models whose query differs only in relation names (574 queries vs 2,997).

Changes

• Session batches share queries as Python's prepare_binding_queries does (stubs, keys, two-member minimum) and re-analyse inexact results alone. Eligibility is never wider than Python's.
• NativeModelAnalysisSession.sharing reports shared and re-analysed counts.
• New model_analysis_cpu_ms phase (process CPU), one timer shared with wall phases.
• Ratio tool --compile-arg; CI ratio entry gates model_analysis_cpu_ms at 1.50x Python + 50ms with --select=*.
• No Python deleted: Python's sharing still serves the python engine.

Changed expectations: model_analysis_cpu_ms in compile JSON timings; new ratio E2E case; parity cases 22/6/2 and 4/0/3 shared/re-analysed/unshared.

Known residuals:

• If a member's stubbed SQL fails to normalise, native analyses it unshared; Python raises for the whole batch. No realistic trigger.
• model_analysis_cpu_ms is process-wide, so without --select=* it includes early-lint threads.
• No test fails if binding_shapes leaves the key: it follows the same relation facts as analysis_shapes, whose removal fails.

Verification

• Model analysis CPU, dense 3000, same machine, 3 alternating runs: 13.0-13.5s before, 4.8-4.9s after, Python 4.2-4.3s. 8 runs: 1.09-1.145x. Sharing off: 3.17x, fails.
• Sabotage: skipping re-analysis or dropping analysis_shapes from the key fails parity.
• Dense output and diagnostics identical to Python. Differential: 321 identical, analysis-session deferrals unchanged.
• Compile E2E; cargo test, clippy, fmt (1.99.0); make check-ci; make test (11949 passed); hygiene.

[kvlonge]

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