initcache.va

// Enhancement-326: an init cache slot was typed by a CROSS-NAMESPACE `Value`
// comparison, so its recorded type depended on a numeric coincidence.
//
// `collapse_implicit` holds INIT-function values; `build_init_cache` tested
// MAIN-function values against it. `Value` is a bare u32 index, so membership
// silently succeeded whenever the two independent value counters collided. On a
// collision an f64 cache slot was recorded as `Type::Bool` -> i8, so the store
// side emitted `trunc double .. to i8` and the noise loader read the slot back as
// a raw i8 straight into `fmul i8 %x, double %y`. That invalid IR reached LLVM,
// where the SHIPPED compiler died with EXC_BAD_ACCESS in DoubleAPFloat::multiply.
//
// The ingredients that make the collision reachable (each verified essential by
// ablation): a noise source whose POWER is an op-independent NON-parameter
// expression (so it lands in a cache slot rather than a Param/Const output), an
// analog operator that creates an implicit equation (`idt`) under a
// parameter-only NON-constant condition (so the collapse flag is a runtime phi,
// not a folded constant), and the noise live in a contribution. The two filler
// statements are load-bearing only in that they position the value counters --
// deleting one shifts an index and breaks the coincidence, which is why the
// original 917-byte fuzz reproducer resisted line-by-line minimization.
`include "disciplines.vams"
module initcache(a, b);
    inout a, b;
    electrical a, b;
    parameter real p = 1.0;
    parameter integer s = 1;
    real n, q, d;
    analog begin
        d = d + V(a, b) * 1.0;
        d = d + V(a, b) * 2.0;
        n = white_noise(abs(p));
        if (s) q = idt(V(a, b));
        I(a, b) <+ n * q + 1e-30 * d;
    end
endmodule