Measured live at Rithwic 2026-08-06 with ACDREAM_DUMP_EDGE_SLIDE=1.
Six branch2/steep-cliffslide events, every one reporting
curN=(-0.954,0.000,0.301) lastN=(-0.954,0.000,0.301) angle=0.0000
apply=False, outcome degenerate-cross/last-known. That is decision-table
row 1 of the research doc, verbatim.
CTransition::cliff_slide @0x0050a6d0 takes its slide direction from
cross(steep contact normal, last_known_contact_plane.N) — it needs the
surface the mover was STANDING ON as the second vector. acdream's
CollisionInfo.SetContactPlane latched the last-known group on every
call, so by the time cliff_slide ran, last-known had already been
overwritten with the steep face itself: the cross product of a vector
with itself, which is zero. Degenerate direction, no slide, walk off
the cliff.
Retail's COLLISIONINFO::set_contact_plane @0x00509d80 is 22 bytes and
writes the CONTACT group only; the last-known group has four writers,
none of them that function. So the four writes are DELETED and a new
InitContactPlane mirrors CTransition::init_contact_plane @0x0050e850,
writing both — the start-of-transition seed, where there is no earlier
surface to remember. Only check_contact's SUCCESS branch calls it. The
other eleven call sites keep the narrowed setter. This is a port, not a
suppression: no guard, no grace period, no flag.
The user's own A/B was the discriminator: Neftet's block plateaus hold
(188 branch3/precipice-slide events, all before the teleport) while
Rithwic's terrain cliff fails (6 branch2 events, all after). I had
predicted the opposite — that terrain would be the flat-normal case —
and position plus timeline corrected me, not reasoning.
NEW DISCRIMINATING TEST, because the suite had none. It was green both
before and after the production change, so nothing in it defended this
behaviour. Issue32LastKnownContactPlaneTests seeds a walkable plane,
asserts a steep mid-transition contact leaves it intact, and asserts the
resulting cross product is non-degenerate. Sabotage-verified: restore
the four writes and both discriminating rows fail while the
InitContactPlane control keeps passing — the pair separates 'the latch
is gone' from 'nothing writes last-known at all'.
Two existing tests corrected rather than deleted.
PhysicsSetPositionTests.FailedCheck_MapsCollisionHandlerResultToRetailError
passed BECAUSE of the latch (the file the research named); its hook now
populates both groups explicitly, since it asserts report plumbing, not
setter semantics. RetailEdgeResponseOrderingTests.TransitionalInsert_
DegenerateCliffSlideOk_ContinuesOuterRetry was predicted to fail and did
not — it now passes for a DIFFERENT reason (last-known absent rather
than clobbered, which retail also answers with OK_TS). Its comment
described the deleted behaviour and is corrected to say so, and to say
it does not discriminate this fix.
Also repairs the #338 probe. Its first placement in
PlayerMovementController printed nothing across 11,523 live log lines —
the wrong one of two resolve call sites — so it moves to
PhysicsEngine.ResolveWithTransition where every caller passes through,
filtered to the player. The dead site is removed rather than left in
place; a probe that never fires is worse than none. The flag test now
precedes the interpolated string: building it eagerly cost 128 B per
resolve with the probe OFF, which Slice I1's zero-allocation gate caught.
Suite 11,234 passed / 4 skipped / 0 failed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
