Stage 0's measurement (previous commit) says the projection is redundant,
so AD-10 retires by deletion rather than by narrowing.
The measurement. With the sample forced to null at BOTH fork sites, from a
clean build:
* a remote running 30 ticks down a 31-degree walkable ramp produces a
BIT-IDENTICAL trajectory, position for position;
* on an 8.4-degree ramp the two differ by at most 2.8e-5 m in Z after 30
ticks (0.03 mm) and are identical in X and Y — float ordering noise
from projecting twice against the same plane rather than once;
* the whole AcDream.Runtime.Tests suite is unchanged.
That is what redundancy looks like, and the arithmetic explains it. The
boundary projection and Transition.AdjustOffset are the same operation
(v -= N * dot(v, N)) against the same plane, and the composition is
idempotent: a vector already on the plane has dot(v, N) == 0, so the
sweep's own projection is a no-op on an already-projected offset and the
full-strength projection on an unprojected one. Either alone produces the
same offset. On terrain a THIRD mechanism, ValidateWalkable's push-out,
re-seats the sphere on the plane every sub-step regardless.
Deleted:
* both RuntimeRemotePhysicsUpdater sample sites (the host and no-host
fork branches carried the block verbatim — the AP-22 shape, a row
naming one site where two exist);
* the terrainNormal parameter and projection block on
RemoteMotionCombiner.ComposeOffset;
* the same block on ComputeOffset, which has no production callers but
held a second copy of the divergence, so leaving it would have made
the row's retirement false;
* PhysicsEngine.SampleTerrainNormal, now callerless.
Removing the parameter rather than passing null is deliberate: it is what
makes a future one-site-only regression a compile error instead of a
silent half-fix.
Two tests went with it —
ComputeOffset_RootMotionFallback_SlopedTerrainNormal_ProjectsZOntoSlope and
its flat-ground twin. Both were weak on their own terms: they drove the
production-dead ComputeOffset and computed their expected values by
re-implementing the projection formula, so they could catch a wrong
MULTIPLY but never a wrong PLANE — which is exactly what the divergence
was. The surviving coverage is geometric and runs the production tick.
Three claims in the old row did not survive contact with the code and are
recorded in the retired row rather than quietly dropped: the justification
(remotes do run the sweep); the description of ComposeOffset's guard as
"interpolation-active" when the code reads `if (!interpolationOverwrote`;
and the roof clause, stale since Bug B gated the sample on OnWalkable —
a steep roof is OnWalkable == false, so the path never ran on #32's
geometry. The retail anchor is corrected too: pc:272296-272346 truncated
both the sliding-normal validity gate at the head and the entire safety
push-out block at the tail. The whole function is 0x0050a370,
pc:272271-272393.
This does not fix#32 and does not partially fix it. #32's remote half was
already closed at 204d0ae0. What deletion does improve is the case #32
never covered: a remote on a WALKABLE non-terrain surface — a bridge, a
dock, a gentle roof, a ramp inside a building — where the terrain sample
returned the plane of the ground far below and applied a wrong plane
rather than none. That surface now gets the body's own committed contact
plane, because that is the only projection left.
The planning contract this work executed is committed alongside as
docs/research/2026-08-06-ad10-contract.md.
Release build 0 errors. Complete solution suite 11,196 passed / 4 skipped
/ 0 failed against the ef976c6d baseline of 11,195 / 4 / 0 — reconciled
exactly as +3 new Runtime tests and -2 deleted Core tests.
Visual gate outstanding: G1 (the ~5 Hz staircase on rolling terrain) is
the veto criterion and runs first; then slope-descent smoothness, a
walkable non-terrain surface, the #32 roof scenario, and flat ground.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Port retail portal viewport projection and reveal behavior, preserve outbound combat style, drive remote and local grounded movement from authored CSequence root frames, and reuse the local prepared pose so animation hooks advance once.
User-verified portal, observer movement, combat stance, and short-tap locomotion gates. Release build passed with 5,767 tests and five intentional skips.
Co-authored-by: OpenAI Codex <codex@openai.com>
The retail movement-manager family the R4 MoveToManager port left as
do-not-invent seams (decomp §9f/§9g). Faithful C# ports of retail's
PositionManager facade + StickyManager + ConstraintManager + the
TargetManager voyeur system, with full conformance tests. NO wiring yet
— purely additive, no behavior change. Wiring (retiring TS-39 sticky +
AP-79 target adapter) is R5-V2/V3.
New Core classes (src/AcDream.Core/Physics/Motion/):
- StickyManager (0x00555400): follow-a-target steering. adjust_offset's
dense x87 mush decoded via ACE (StickyRadius 0.3, StickyTime 1.0,
follow speed ×5 / fallback 15) — speed-clamped signed-distance steer +
bounded turn-to-face; 1 s watchdog; Ok→initialized / non-Ok→teardown.
- ConstraintManager (0x00556090): the server-position rubber-band leash.
90% IsFullyConstrained jump gate + grounded linear brake taper.
Structural only — acdream never ARMS it (retail arms from
SmartBox::HandleReceivedPosition, which acdream lacks, with two x87
constants BN elided). IsFullyConstrained stays false = TS-35 behavior;
leash-arming + the unknown constants are a deferred issue.
- PositionManager facade (0x00555160): lazy Sticky/Constraint + fan-out.
- TargetManager (0x0051a370) + TargettedVoyeurInfo: the peer-to-peer
voyeur subscription system (0.5 s throttle, 10 s staleness,
send-on-drift-past-radius, dead-reckon GetInterpolatedPosition). A
faithful superset of the AP-79 adapter — SetTarget subscribes ON the
target; the target's HandleTargetting pushes updates back.
- IPhysicsObjHost: the CPhysicsObj back-pointer seam (position/velocity/
radius/contact/GetObjectA + target-tracking fan-out) the App wires per
entity in V2/V3. MotionDeltaFrame: mutable retail-Frame delta accumulator.
Supporting:
- TargetInfo extended to the full retail 10-field struct (additive
defaults keep the R4 4-arg call sites compiling).
- MoveToMath: signed CylinderDistanceNoZ, NormalizeCheckSmall,
GlobalToLocalVec.
- Rename: the misnamed AcDream.Core.Physics.PositionManager (a remote
anim+interp per-frame combiner, NOT the retail facade) → RemoteMotion
Combiner, freeing the name and removing the ambiguity that breaks every
file importing both Physics + Physics.Motion (GameWindow will in V2/V3).
Tests: 42 new conformance cases (Sticky/Constraint/Position facade +
TargetManager incl. the full cross-entity voyeur round-trip). Full suite
4006 green (+2 skipped), no regressions.
Decomp + ACE cross-ref + port plan: docs/research/2026-07-03-r5-managers/.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-03 19:34:49 +02:00
Renamed from tests/AcDream.Core.Tests/Physics/PositionManagerTests.cs (Browse further)
