The diagnostics refactor (72c0109c) kept the debug HUD and the simulation on one
fire-control model by routing HitProbability THROUGH Inspect. That was the right
invariant implemented at the wrong level: every AI and turret evaluation, per
weapon per tick, built the full diagnostic struct and ran IsRevealed's sort, the
aim-point cell walk and HullKernel's allocations and per-cell exp -- and only
then checked whether the target was even in range. A Soul pass measured 9.87 us
and 448 B per call on paths that used to bail in nanoseconds. Operator: it made
the hot path look like a diagnostic one.
One model is now shared at the level of its factors. HitProbability runs its
cheap gates first and bails, exactly as it did before the refactor, and computes
PSensor, PSpread and POnHull only for a shot that is actually possible. Inspect
calls the same four factor functions and fills the struct unconditionally,
because a HUD wants to see why a shot is impossible. The HUD panel and its one
test consumer are unchanged.
Two tests make the design enforced rather than conventional:
- HitProbabilityMatchesInspect: exact equality across 21 configurations covering
every gate the hot path bails on plus Precision and info sweeps. Dies when a
factor changes in only one caller.
- GatedOutHitProbabilityAllocatesNothing: 1000 out-of-range calls allocate zero
bytes. Dies when HitProbability is routed back through Inspect.
Both mutations verified RED and distinguished from a compile failure, with
restores checked against a hash taken before mutating.
231 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
9.1, blocking. FireControl seeded CultMath's xorshift32 per shot and drew once.
NextFloat returns the top 24 bits of the FIRST round, and one round cannot carry
ShotId's low-bit variation into the high bits -- so every zone's hit roll spanned
about an eighth, with the threshold fixed at zone creation. A shot was a step
function of p: below the band it never hit, above it always did. An unaided
shooter hit 0 of 18,926 shots. That is the operator's "no hits registering," and
being stable per zone is why it came and went.
The seed derivation is unchanged -- still a pure function of (zone, shot id),
Cut 6b's rule -- and only its diffusion into the first output changes, via an
fmix32 finalizer. Measured across six zone seeds: hit fraction at p=.3 went from
0.0% in five of six zones to 29-31%, and at p=.6 to 58-61%, with the roll
spanning the full range.
The defect survived three cuts because nothing in the campaign measured a
distribution -- every fixture uses accuracy 1 or 0, and a constant satisfies a
determinism test perfectly. TheDieIsUniform now asserts the observed hit
fraction is within .08 of p across several zone seeds, which is the test whose
absence was the real defect.
9.2. Sigma is floored at half a cell. HullKernel approximates a continuous
Gaussian with a discrete sum over occupied cells, and that sum collapses once
sigma falls well under one cell against an aim point that is not exactly on a
cell centre -- true of every shipped hull. Past Precision ~4, pOnHull went to
zero and a better targeting system made the weapon unable to hit anything.
9.3. An active entity's Target is now always a live entity it has info on.
Cut 8's _active guard could never have fixed the operator's LockWeapon crash --
that code only runs on active entities, and cut8's own harness header says so.
The live route was dock, target dies, undock; Deactivate disposed the
subscription that nulls a stale Target and Activate reseeded EntityInfoGathered
without reconciling it. Still not a TryGetValue in LockWeapon.
229 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>