FixedPointItemResolution now throws "the model never meets a cycle" if a pool
is still open after its fixed point settles, instead of silently
force-settling in the reverse of first-encountered order. The sweep already
gates its own draws (LanesReaching) so a genuine cross-item cycle is never
actually presented to the model; a thrown exception here means the gating
missed a real case, not a numbers question the model should answer with its
own competing tie-break. The stale "opposite tie-break" comment describing
the deleted behaviour is gone with it.
Three of Soul's five probes (soul125-FireControlCut123Tests.probe.cs) are
committed here, ported from reflection onto the public Fire/Commit/Apply path
this file's own tests already use -- FireCentre/FireUntilHit retries until
the committed hit lands on the expected column, the same idiom
ZeroDepositDoesNotStall and PooledPathHonoursTheItemThresholdLikeTheSoloPath
already use, never a direct call into the private Apply:
- ZeroDepositContributorFiresNoItemDamageEvent (P3): a zero-deposit
contributor must not fire a phantom ItemDamage event. Kills Entity.cs's
`incoming[i] > 0f` boundary flipped to `>=`, and a swap that reports the
post-split share instead of the incoming one (F2).
- ClippedLaneDoesNotHoldThePool (P4): an item beyond every lane's own
penetration clip must never be waited on.
- OwnRepeatVisitDoesNotHoldThePool (P5): a lane's own later visit to an item
must not hold that item's pool open -- only other lanes are future
reachers. Kills FireControl.cs's `alreadyContributed` guard being deleted
(F1).
The other two probes (Soul5_AngledMirrorAndConservationWithSharedItems,
Soul5_AngledMirrorCycleShapes) are NOT committed. They drive Apply directly
through reflection with a hand-built PendingShot.Outcome so they can sweep
1400 exact, mirrored (bearing, lateral) pairs per test and compare against an
independent model -- Commit's own bearing is reachable through the public
path (shooter/target positioning), but its lateral is a random draw from the
per-shot RNG, not a value either test can set or predict without either
control of that RNG's seed and Silhouette's exact sigma, or a decoy Apply
entry point. Reaching them without reflection needs one of:
(a) `[InternalsVisibleTo]` for Aetheria.Shared.Tests on FireControl.Apply, or
(b) a public/internal test seam that takes a pre-built ShotOutcome (bearing,
lateral) and runs Apply's own walk against it.
Neither is authorized by this fix batch's map; this is reported rather than
silently reflected. The Soul5_Alloc benchmark (also reflection-based, and not
one of P1-P5) is likewise left uncommitted.
Mutation verification (docs/fire-control-cut.md, Cut 12.3, one-apply-path):
F1 and both F2 mutants confirmed killed by the new tests; the earlier
fix batch's own forced-sequential, resolve-one-pool-only, and cycle-throws
mutants (Soul's M3/M4/M6 family) still die against CycleTieBreakIsMirrorSymmetric
and friends.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cccf3a2ac4