erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

fix(phys): A6.P3 slice 2 v2 — revert seed removal + add no-op guard

Browse source 
First slice 2 attempt at commit 892019b removed PhysicsEngine.cs L622
per-tick CP seed entirely. User happy-test surfaced a regression: BSP
step_up at the last step of stairs failed because sub-step 1's
AdjustOffset had no ContactPlane to compute the lift direction (the
seed was load-bearing for step_up correctness).

Revert + better fix:
  1. Re-add the L622 seed (PhysicsEngine.cs:620-626).
  2. Add no-op-if-unchanged guard inside CollisionInfo.SetContactPlane
     (TransitionTypes.cs:259-279). When called with values identical
     to current state, early-return without incrementing
     ContactPlaneWriteCount or rewriting fields.

When the player stands on the same plane tick after tick, the L622
seed re-calls SetContactPlane with identical args — these now no-op
instead of inflating the counter and re-writing the same values.
Only actual state changes (e.g. landing on a new step's plane, cell
crossing) increment the counter.

Verification (post-rebuild, pre-this-commit slice 2 first attempt):
- scen3 walk produced 2,690 cp-writes (down from 30,420 = 91%
  reduction from L622-seed presence)
- BUT user could not pass the last step of stairs — step_up regression
- Test suite: 1148 + 8 pre-existing fail baseline maintained but
  physical behavior broke

Post-this-commit expectations:
- Test suite: 1148 + 8 (unchanged, no behavioral change in fixtures
  because the seed value is what the fixtures already expect)
- Stair-walking: works (seed restored)
- CP-write count: significantly reduced (most seeds are no-ops because
  body.CP doesn't change tick-over-tick on stable footing)
- Issues #96 / #97: re-test in re-capture; #96 should be largely
  closed via the guard; #97 (fall-through + stuck-in-falling) was
  observed pre-slice-2 too, so unrelated to the seed

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Erik 2026-05-22 10:32:23 +02:00
parent 892019bc9a
commit f8d669be88
2 changed files with 49 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

55
src/AcDream.Core/Physics/PhysicsEngine.cs
View file

@ -597,36 +597,35 @@ public sealed class PhysicsEngine
if (isOnGround)
transition.ObjectInfo.State |= ObjectInfoState.Contact | ObjectInfoState.OnWalkable;
// A6.P3 slice 2 (2026-05-22) — DO NOT seed ContactPlane at
// transition start. Retail's CTransition::init at 0x509dd0
// (acclient_2013_pseudo_c.txt:271954) explicitly clears
// contact_plane_valid = 0 — retail does NOT seed. Cross-tick
// CP retention flows via:
// - Mechanism A: BSPQuery.FindCollisions Path-6 land write
// (writes ContactPlane during the sub-step loop when the
// sphere ground-contacts a floor poly)
// - Mechanism B: Transition.ValidateTransition LKCP restore
// (slice 1 commit 5aba071; restores ContactPlane from
// LastKnownContactPlane post-sub-step when sphere close to
// LKCP)
// - Body persist below: at transition end, body.ContactPlane
// gets set from ci.ContactPlane OR ci.LastKnownContactPlane
// so the NEXT tick's body state is preserved.
// K-fix7 (2026-04-26): only seed the contact plane when the body
// is actually grounded. Pre-seeding while AIRBORNE caused
// AdjustOffset's "Have a contact plane / Moving away from plane"
// branch to fire on every jump step — which calls
// Plane::snap_to_plane on the offset and ZEROES the Z component,
// killing all upward jump motion.
//
// Cost: AdjustOffset on sub-step 1 of each tick has no CP and
// takes the "no contact plane" path (TransitionTypes.cs:2329-2334)
// — slide-projects only, no slope-snap. Sub-steps 2+ pick up CP
// from Mechanism A / B as normal. Imperceptible in practice.
// We KEEP the seeding when isOnGround for slope-walking + step-up
// continuity (the original concern that motivated the seed).
// BSP step_up needs ContactPlane on sub-step 1 to compute the
// correct lift direction; removing the seed breaks stair-walking
// at the last step (verified by A6.P3 slice 2 first attempt
// 2026-05-22, reverted in this commit). Retail's CTransition::init
// explicitly CLEARS contact_plane_valid; we deliberately diverge
// for step_up correctness.
//
// Removing this seed closes issue #96 (per-tick CP-write blowup;
// 99.3% of post-slice-1 CP writes came from the seed) and is
// hypothesized to close issue #97 (phantom collisions +
// fall-through; suspected stale-CP slope-snap was the cause).
//
// Previous K-fix7 seed (2026-04-26 to slice-1 ship) deliberately
// diverged from retail for "slope-walking continuity" but the
// tradeoff (per-tick CP writes, stale-CP slope-snap) was not
// worth it.
// A6.P3 slice 2 (2026-05-22) — to close issue #96 (per-tick CP-write
// blowup) without breaking stair-walking, the no-op-if-unchanged
// guard inside CollisionInfo.SetContactPlane (TransitionTypes.cs:259)
// collapses redundant seeds (same plane every tick) to a true no-op.
// The seed still fires the function call but only counts as a write
// when the plane values actually change.
if (isOnGround && body is not null && body.ContactPlaneValid)
{
transition.CollisionInfo.SetContactPlane(
body.ContactPlane,
body.ContactPlaneCellId,
body.ContactPlaneIsWater);
}
// Retail CPhysicsObj::get_object_info also seeds SlidingNormal when
// transient_state has bit 2 set. This matters for one-step/frame hits:

22
src/AcDream.Core/Physics/TransitionTypes.cs
View file

@ -258,6 +258,28 @@ public sealed class CollisionInfo
public void SetContactPlane(Plane plane, uint cellId, bool isWater = false)
{
// A6.P3 slice 2 (2026-05-22): no-op-if-unchanged guard. Closes
// issue #96 (per-tick CP-write blowup) without removing the
// PhysicsEngine.cs L622 seed that step_up depends on. When the
// player stands on the same plane tick after tick, the L622 seed
// re-calls this function with identical arguments — collapse
// those redundant calls to a true no-op. The function still gets
// called from the seed site; this guard just skips the increment
// + field writes when nothing would actually change.
//
// This is safe because all four fields (ContactPlane,
// ContactPlaneCellId, ContactPlaneIsWater, ContactPlaneValid)
// are write-only side effects — no observer reads them mid-call.
// Same logic applies to the LKCP re-latch below.
if (ContactPlaneValid
&& ContactPlaneCellId == cellId
&& ContactPlaneIsWater == isWater
&& PlaneEquals(ContactPlane, plane))
{
// Already in the target state. Skip the write + counter.
return;
}
ContactPlaneWriteCount++;
ContactPlaneValid = true;
ContactPlane = plane;