Restore stochastic transition() semantics, schedule-independent $rdist_normal, and measured-law cross acceptance

[BucketSran]

Summary

This series completes the stochastic-parity work started in 06526fa and continued in 28c9100, fixing the remaining semantic regressions on the python engine, making cross-engine random sequences schedule-independent, and replacing the cross-acceptance heuristic with a law measured from strict-mode Spectre.

1. Restore transition() semantics for stochastic targets — both engines

06526fa bypassed the transition() ramp for $rdist_normal-derived targets in both engines (python codegen direct-write + rust snap). 28c9100 fixed the rust side but left the python bypass, so as of upstream main the packaged default engine still snaps and the engines disagree with each other (measured on the vaBench dither tb gold: python 252.2 ps vs stored Spectre; python-vs-rust 275 ps).

This PR removes the python codegen bypass and the remaining rust snap path (force_immediate_random_target), re-enables active-ramp completion breakpoints (deterministic, no prediction involved), and keeps the predictive target-change breakpoint skip with 28c9100's finer gating. Result: dither tb gold 3.25 ps vs stored Spectre on BOTH engines, python == rust bit-exact.

2. Hash $rdist_normal on per-seed draw index instead of wall-clock time

The time-hashed RNG made every draw bit-sensitive to the exact event time: femtosecond-level schedule differences between engines produced entirely different sequences once real ramps were restored. Hashing (seed_bits, per-seed draw index) — identically in python (per-instance counters) and rust (thread-local counters, reset per trace) — makes sequences schedule-independent and matches the LRM's sequential per-seed stream model. The noise_gen parity test is un-xfailed and passes bit-exact over 1000 periodic draws; the bundled example is rewritten to the recommended @(timer(dt, dt)) sample-and-hold form.

3. Cross-acceptance window driven by the measured Spectre law

A 52-measurement DOE against strict-mode Spectre ($abstime probe DUTs, slopes over 2 decades, irrational-flavored roots) shows Spectre accepts a cross event late by

Delta = 0.5 * reltol * |V_cross| / |slope|

independent of maxstep (1p–500p), errpreset, and explicit cross time_tol. The law retrodicts the April R014 residual exactly (0.5 × 1e-4 × 30 ns = 1.5000 ps vs 1.500025 ps measured); the April "maxstep=1p removes the delay" observation was a grid-aliasing artifact (root aligned to the step grid).

Accordingly:

• the 0.25 × transition-ramp window is replaced by the law (kappa = user_factor × 0.5 × reltol folded in the netlist runner; rust computes slack = kappa × |V_root| / |slope| with the magnitude evaluated at the interpolated root);
• the factor is an explicit opt-in via simulatorOptions options evas_cross_acceptance_slack_factor=1.0 and defaults to 0.0 in all presets (per the 2026-04 closure decision, exact/analytic stays the default and benchmark flows must not enable tolerance-compatible behavior);
• accepted-event-time is now pre-phase only: post-phase crossings keep exact interpolated timing (previously they were moved to an arbitrary accepted grid step without a matching breakpoint — 10–18 ps error observed).

Validated end-to-end: with factor=1.0 EVAS reproduces the measured strict-Spectre lateness on held-out external-ramp probes to 0.001–0.05 fs.

Verification

check	result
engine/compiler/netlist	523 passed (incl. both 28c9100 tests, adapted)
examples	12 passed
rust parity	16/16 (noise_gen un-xfailed, bit-exact)
vaBench dither tb gold	3.25 ps vs stored Spectre, both engines, python == rust bit-exact
37-entry release-evidence sweep	exactly 1 signal changed vs pre-fix baseline (the fixed one)
law-mode probes vs measured Spectre	0.001–0.05 fs

Notes

• Draw sequences change (second deliberate change after 06526fa); waveform baselines capturing $rdist_normal values must be regenerated.
• Known v1 scope limit: phase classification treats cross exprs referencing contributed branch nodes (including rails used as contribution node2) as post-phase, so law mode currently applies to ground-referenced cross exprs on source-driven nets. Refining this classification is a follow-up.
• Full evidence: vaEvas/testspace/edge-timing-triage-20260611/ and vaEvas/testspace/cross-lateness-doe-20260612/.

🤖 Generated with Claude Code

[BucketSran]

Added 0ef7004 — phase classifier refinement: only the driven node (node1) of a V() contribution counts as contributed; node2 is a reference. Previously every cross expression referencing a rail used as a contribution node2 (the common V(in, VSS) style) was classified post-phase, silently excluding it from pre-phase-only features. With the fix, V(in, VSS)-style probes reproduce the measured Spectre lateness in law mode to 0.001–0.05 fs. Default behavior unchanged: full suites green and a 37-entry release-evidence sweep shows zero signal-level diffs. Guard: TestCrossPhaseClassification.

🤖 Generated with Claude Code