[kvlonge]

Why

Custom rules ran 1.2-1.3x slower (wall) under native-preview than python on the dense 3,000-model project. In CPU the gap was real: 1.151x (1794 vs 1559 ms, 5 alternating runs). Native session analyses built a new InferredColumn per model column, while Python's compact analysis shares one object per distinct column. The custom-rule host payload therefore pickled more distinct objects (64.0 MB vs 59.6 MB) and hosts decoded more.

Changes

• NativeModelAnalysis keeps one InferredColumn per distinct column row for the session. Values are unchanged; only object identity is shared. Host payload is now 59.6 MB.
• New custom_rules_cpu_ms: CPU of the custom-rule thread plus finished child processes (the rule hosts, via os.times; children are 0 on Windows).
• Ratio tool reports it. New CI ratio entry gates custom_rules_cpu_ms at 1.25x python + 50 ms. The margin is wide because CI noise for this phase is unproven and the rule hosts share cores; the deterministic column-object count in the parity test is what pins the interning, so the gate only catches gross regressions.
• Stacked on perf(native): share equal model analyses in the native session and guard their CPU #1073 for the ratio tool's engine and phase options.

Changed expectations, each new:

• custom_rules_cpu_ms in compile JSON timing keys and in the dense cold E2E (> 0).
• Ratio E2E case for the custom-rule phase.
• Analysis parity: native column objects equal distinct column values (12 in the shared case; generated case asserts equality).

Verification

• custom_rules_cpu_ms ratio, dense 3000, same machine: 1.151 before; after 1.007, 0.998, 0.996, 1.003, 1.005 (5 comparisons). The 1.144 sabotage is caught by the parity count, not this gate.
• Sabotage, no sharing: ratio gate fails (1.144); parity fails (1863 objects vs 903 values; 54 vs 12).
• Differential: 321 identical; deferrals unchanged.
• Rule engine unit, ratio unit/E2E, compile JSON, analysis-session parity, seams and dense cold E2E tests; make check-ci; hygiene.

[kvlonge]

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