erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 2
acdream/docs/plans/2026-07-29-physics-parity-campaign.md

23 KiB
Raw Blame History

Campaign P — Physics Retail-Feel Parity

Filed: 2026-07-29. Directed by the user as a pre-vendor-management detour after the same-day physics/collision retail-fidelity audit. The world-interaction program (Slice 5, vendor browsing) resumes when this campaign closes.

Execution model: Claude drives autonomously slice-to-slice. Sonnet subagents implement bounded chunks against this plan's specs; an Opus review subagent gates every slice boundary. The only user stops are (a) the final batched connected visual matrix, (b) a DO-NOT-RETRY conflict, (c) anything destructive. All the CLAUDE.md workflow rules apply: grep-named-first → pseudocode → port → conformance test; register moves in the same commit; no workarounds.

The parity goal (the autonomy contract)

"Retail Movement Parity v1" — the campaign is DONE when all of the following are auditable-true:

  1. Zero physics-scope temporary stopgaps. TS-1, TS-4, TS-5, TS-23, TS-46 retired by porting the retail mechanism. TS-24, TS-35, TS-40 either retired or re-classified (IA/AD) with a recorded justification.
  2. No feel-affecting approximations left unargued. AP-7 resolved by decoding retail's friction state gate; AP-10 restored to retail's 0.1 m water sink-in; AP-25 replaced by the retail effective-skill chain (vitae/enchantment-aware); AP-71 (check_entry_restrictions) ported.
  3. Issue ledger: #262, #165, #166, #116, #167, #72, #153 closed; the stale "pending visual gate" statuses on #172/#173/#174/#175/#41 reconciled (folded into the final matrix below). Explicitly excluded: #235 (user-deferred 2026-07-27) and #256/#257 (lifecycle/memory, not physics feel — separate track).
  4. Verification: each port carries decomp citations + conformance tests; dotnet build + full Release suite green at every slice commit; one batched connected visual matrix (below) passes at the end, run by the user.

Anything not in this list is out of scope for the campaign — file it, don't chase it.

Slices

P1 — Stat-coupled movement (burden / stamina / vitae) — retires TS-5, AP-25, the burden gap

Status (2026-07-30): COMPLETE. Landed at 9355ddce; Opus review APPROVE at 001e466d (which also retired UN-8 — the CanJump polarity is byte-PROVEN load < 2.0 from the PDB-paired binary — and recorded the PK-timer jump-cost decode for P3). Full Release suite 9,880/0/5 at the slice gate. Retail's vitae/enchant chain reuses the M3 bucket-4 representation; JumpStaminaCost never refuses (weak-jump only) — the plan's formula shorthand had the +0.5 operand wrong and the implementation follows the decomp's (load+0.5)*power*8+2. AP-127 was filed for the then-bounded bonus properties and retired by #268 on 2026-07-31.

Today PlayerWeenie.SetBurden has zero callers, CanJump is always true, JumpStaminaCost is 0, and pushed run/jump skill is attributeBonus + init + ranks only. Retail modulates movement by character state continuously.

Retail anchors (named decomp, verified 2026-07-29):

  • ACCWeenieObject::{CanJump 0x0058c400, JumpStaminaCost 0x0058c440, InqJumpVelocity 0x0058c520, InqRunRate 0x0058c560, InqMaxRunRate 0x0058c5a0} — thin delegations to the qualities DB (pseudo-C ~406512).
  • CACQualities::{InqMaxRunRate 0x00591b20, CanJump 0x00591b50, JumpStaminaCost 0x00591b90, InqRunRate 0x00592800, InqJumpVelocity 0x00592980, InqLoad 0x0058f130} (pseudo-C ~412901–413975, ~409756) — the real load/skill/vitae composition. InqJumpVelocity ends in sqrt(GetJumpHeight(...) * 19.6) (pc 413975).
  • MovementSystem::{GetRunRate 0x006b0950, GetJumpHeight 0x006b09b0, JumpStaminaCost 0x006b0a40} (pseudo-C ~695958) — GetJumpHeight readable: LoadMod(load) * (skill/(skill+1300) * 22.2 + 0.05) * power / scaling, floor 0.35; JumpStaminaCost readable: ceil(((power + 0.5) * load) * 8 + 2) on the arg3==0 branch.
  • EncumbranceSystem::{EncumbranceCapacity 0x004fcc00, Load 0x004fcc40, LoadMod 0x004fcc70} (pseudo-C ~256393).
  • Cross-refs: ACE MovementSystem/EncumbranceSystem C# ports; holtburger if it models load.

Work:

  1. Port the full chain into Core (EncumbranceSystem + MovementSystem statics; PlayerWeenie becomes the CACQualities-shaped composition). Where BN x87 mush blocks a branch (GetRunRate body), use Ghidra MCP or the ACE port as the tiebreaker and cite which.
  2. Determine, from CACQualities::InqRunRate's own body, exactly which skill level retail feeds (base vs enchantment/vitae-adjusted) and port THAT — scoped to the run/jump query path only, reading vitae + relevant skill enchantments from the M3 active-effect state. Do not build a general effective-skill engine.
  3. Plumb the inputs from Runtime: burden (EncumbranceVal/capacity from PlayerDescription + property updates), current stamina (vitals), vitae. Extend RuntimeMovementSkillState (J4.4 seam) so updates flow mid-session, same as run/jump skill today.
  4. Wire the existing jump_is_allowed stamina-refusal branch and the ReportExhaustion dual-dispatch gate (R3-W4 seam) to a real consumer, matching retail's refusal/weak-jump behavior.
  5. Conformance tests: formula tables (golden values incl. 800-skill cap, load-mod knees at 100%/200%, stamina cost ceil), gating tests (no-stamina jump refusal), plumbing tests (burden/stamina/vitae changes move the produced rate). Register: delete TS-5 + AP-25 rows, note the retirement in the same commit.

P2 — Response-layer edge family — retires TS-1, TS-4, AP-7; closes #166, #116

Status (2026-07-30, FINAL — Campaign P final physics slice): TS-1 and AP-7 retired same-day as originally recorded. TS-4 is now ALSO RETIRED — the oracle follow-up pass (docs/research/2026-07-30-ts4-116-oracle-plan.md) found the freeze the first implementation attempt hit was one layer downstream of Path 6 (inside AdjustOffset's crease projection against a purely-vertical offset — a genuine retail-identical degeneracy, not a bug) and ran the plan's own decisive confirming test: Ts4SteepRoofWedgeCaptureTests's horizontal-velocity variant (matching the realistic live-play input that originally validated the shortcut) converges cleanly with the shortcut removed. The Path-6 steep-poly shortcut is deleted from both BSPQuery.cs and FlatBspQuery.cs; the pure-vertical degenerate case is pinned (not fixed) as register row AD-56. #116 is narrowed, not closed: shape-2 (D4 first-airborne-frame hard-stop) is CLOSED — the oracle plan's structural dispatch-routing hypothesis was confirmed by instrumentation with no cdb session needed, and the D4 pin is un-skipped. Shape-1 (tick-22760 lateral-slide loss) got a real, independently-decomp- confirmed fix (Path 6's foot-clear/head-hit branch now returns Collided

  • SetCollisionNormal directly, matching pc:323824-323834/ACE BSPTree.cs:221-230), but the confirming replay showed this does NOT explain tick-22760 itself — that mover is grounded (dispatches through Path 5, not Path 6) and the actual "no normal recorded" mechanism (SpherePath.PrecipiceSlide's find_crossed_edge-false fallback) is independently confirmed byte-exact retail behavior too. The remaining divergence is most likely this test's simplified door-registration fixture, not the response layer — see ISSUES.md #116 and the oracle plan's Addendum 2 for the full trace and the concrete next step (re-run against the faithful Setup-based door registration). AD-55 (the sled slope-flatness constant, split out of AP-7's retirement) is ALSO retired this same slice — byte-proven cos(10°) per the oracle plan's Addendum.

The collision response layer (what happens after a hit): ground friction, cliff edges, downhill landings, near-perpendicular wall slides. One oracle-driven pass; the physics digest's DO-NOT-RETRY table binds every subagent here.

Work (research doc FIRST, then port):

  1. AP-7: decode the state gate on retail's friction block (calc_friction region, pseudo-C ~276702-276705) that lets retail use threshold 0.25 without hammering normal locomotion (the reverted L.3c attempt). Ghidra MCP for the x87 branch if BN is garbled.
  2. TS-1: port the EdgeSlide → PrecipiceSlide / CliffSlide chain (precipice context, steep-plane bookkeeping) replacing our stop-at-edge.
  3. #166: port the landing "sled" (Sledding state set/clear sites; the sled friction constants already sit in calc_friction).
  4. TS-4: replace the Path-6 steep-poly in-place-slide shortcut with retail's SetCollide → Path-4 → ContactPlane landing chain, and remove the two BSP-layer SetSlidingNormal writes (retail's only in-transition writer is validate_transition).
  5. #116: the near-perpendicular lateral-slide loss + first-airborne- frame divergence, driven by the existing tick-22760 replay and D4 pins.
  6. Apparatus: extend the trajectory-replay tests; capture fixtures before changing behavior. Register: delete TS-1/TS-4/AP-7 rows.

P3 — Remote-object residuals — retires TS-46, TS-23; closes #165; narrows/retires AD-25

  1. TS-46: pass the Setup's verbatim sphere LIST into the transition (CPhysicsObj::transition 0x00512dc0 → init_sphere) instead of the two-scalar reconstruction, for local player and remotes; derive remote step-up/step-down from the Setup instead of the pinned 0.4 m. Captured-fixture replays must stay green or be re-baselined with evidence.
  2. AD-25: align the remote post-resolve with the ported handle_all_collisions (grounded-bounce rule) as the player half already did in #182.
  3. #165: remotes visibly penetrate walls before stopping — diagnose against the (now Setup-true) sweep; suspect list starts at the catch-up step length vs sweep sub-steps.
  4. TS-23: parse PlayerKillerStatus from PlayerDescription/property updates and plumb PK/PKLite/Impenetrable onto local + remote player movers (OBJECTINFO::init 0x0050cf30 state bits). Non-PK ACE behavior must be provably unchanged.

P4 — World specials — retires AP-71, AP-10

Status (2026-07-30): COMPLETE, including a same-day Opus review fix. AP-71 landed at d6c3f865 (20 new conformance tests); AP-10 landed at cc8d57a2 (12 new conformance tests). Complete solution suite at that gate: 9,946 total across 9 test projects, 9,941 passed, 5 skipped, 0 failed on a clean run. One run in the same session saw a single unrelated flake (AcDream.Content.Tests.Vfx.RetailDatLoaderTests .AnimationCache_CoalescesSameDidAndAllowsUnrelatedReadsInParallel, a parallel-cache-coalescing timing test untouched by either commit) that passed 3/3 in isolation and on the immediate re-run — full-suite parallel contention, not a regression (independently fixed afterward at dc0468cc).

P4 review verdict: FIX-FIRST (2026-07-30). RestrictionObjPrevalenceInspectionTests (3b5e0992) measured the installed cell DAT: 103,766 of 729,888 EnvCells across 1,293 landblocks — the entire housing estate, restrictionObj GUIDs 0x70xxxxxx — carry a baked RestrictionObj. AP-71's fail-closed default (landed with CanMoveInto deliberately unmodeled, per the original AP-129 row) would have locked every apartment/cottage/villa interior for every player, including its own owner — a live regression, not the "inert in dev content" the row assumed. Fixed at 7a0f836a: ACCWeenieObject::CanMoveInto (0x0058da40) and RestrictionDB::IsAllowedIn (0x005ae8f0) are now ported verbatim, fed end-to-end from CreateObject's HouseOwner/HouseRestrictions/ Monarch PWD-tail fields (previously parsed-and-discarded) plus a new live House_UpdateRestrictions (0x0248) parser, and resolved through a new PhysicsEngine.Objects property wired to the canonical ClientObjectTable in RuntimeEntityObjectLifetime (production fix, not just gate logic — an unwired table still fails closed). AP-129 is narrowed (not retired) to the genuine residual: no sequence-based staleness rejection for House_UpdateRestrictions (low-probability, self-correcting), and the outdoor CLandCell restriction path (a separate DAT structure) remains unported, unaffected by this fix. Gate: AcDream.Core.Tests 4,049 passed / 2 skipped / 0 failed; AcDream.Core.Net.Tests 761 passed / 0 skipped / 0 failed; complete solution suite 9,961 total, 9,956 passed, 5 skipped, 0 failed.

  1. AP-71: port the CObjCell::check_entry_restrictions gate at the head of find_env_collisions (pc:309576) — barred house cells block at the threshold client-side. Landed: the gate is wired at the top of the indoor branch of Transition.FindEnvCollisions; CellPhysics .RestrictionObj is fed from the DAT-baked EnvCell.RestrictionObj field (§4.3's open question resolved via ACE's DatLoader + an independent Chorizite.DatReaderWriter reflection probe — it's a plain per-cell DAT field, not a live wire override) in both the dev and production caching paths, at zero bake-format cost. The mover's CanBypassMoveRestrictions (BF_ADMIN & BF_IMMUNE_CELL_RESTRICTIONS) is decoded via the TS-23 PWD-bitfield pipeline. The original landing deliberately left CanMoveInto unmodeled (fail-closed default, filed as AP-129) — the P4 review found this fails closed for the ENTIRE housing estate and required the fix-first pass described above.
  2. AP-10: restore retail's 0.1 m water sink-in; while there, verify the water-contact step behavior (WATER_CONTACT_TS consumers) against retail and file anything found. Landed: the dry-corner constant is restored (full suite green — the sticky Contact/OnWalkable bit argument held); WaterContact is now produced at every Contact/OnWalkable commit site. No confirmed retail consumer of WATER_CONTACT_TS was found this pass; filed as #264 along with two other explicitly-unverified water items (the ENTIRELY_WATER ethereal/swim terrain-collision exemption, and jump/swim movement-effects) — none block this port.

P4 review addendum (2026-07-30): APPROVED after one FIX-FIRST round. The initial AP-71 landing failed closed with CanMoveInto unmodeled; the prevalence inspection (3b5e0992) proved that locks all 103,766 housing EnvCells. 7a0f836a ports CanMoveInto/IsAllowedIn verbatim (owner/self/null-db admit; unresolved object blocks), captures the previously-discarded HouseOwner/Monarch PWD fields, parses House_UpdateRestrictions 0x0248 live, and wires the canonical object table into the gate. AP-129 narrowed to the sequence-byte and outdoor RestrictionTables residuals. Suite 9,956/0/5.

P5 — Deferred fidelity — closes #167, #153, #72

Status (2026-07-30): item 1 (#167) COMPLETE. Both blockers resolved without Ghidra/cdb — the two x87-elided constants were byte-decoded straight from the matching binary's raw machine code (docs/research/2026-07-30-constraint-leash-constants.md). The leash is now armed at every current acdream inbound-position acceptance seam (ConstraintDistance, LiveEntityNetworkUpdateController, PlayerMovementController.SetPosition/BlipPosition), the per-tick PhysicsBody.IsFullyConstrained push replaces the always-false stub, and register row TS-35 is deleted. Full Core/Runtime/App suites pass with new conformance tests (leash-armed jump refusal, teleport-vs-blip anchor/teardown, taper reduction, remote-tick push). Items 2 (#153) and 3 (#72) remain open.

  1. #167: decode the two unknown x87 ConstraintManager constants (Ghidra) and port leash arming.
  2. #153: the far-teleport arrival onto an unstreamed landblock near a 192 m edge — apparatus first (the issue's own trigger table), then the streaming-gap hold shape ALREADY sketched there (freeze the per-tick resolve until the landblock loads — the async equivalent of retail's synchronous load; this is an AD row, not a workaround, and gets filed as one).
  3. #72: close on the R6 evidence (DAT-authored omega ±1.5 rad/s is live; the cdb confirmation ask is obsolete).

P6 — #262 login run-on-the-spot (live defect)

Probe-instrumented fresh-process login repros (ACDREAM_PROBE_RESOLVE=1

  • net probes) against local ACE; the issue's hypothesis list is the script. Root cause, fix, regression test. No workarounds (no auto-recall, no synthetic position kick). Runs serialized (owns the build tree + client).

P7 — Ledger + camera feel

  1. Retire the stale TS-25 row (outbound stance ships via RawState.CurrentStyle since #219) and refresh TS-24/TS-35/TS-40 classifications.
  2. #115 camera-drag: investigation-only against CameraManager constants (AD-37's vector-nlerp vs retail quaternion-slerp is the prime suspect); fix if a concrete divergence falls out, otherwise re-classify with evidence.
  3. Reconcile #172/#173/#174/#175/#41 statuses via the final matrix.

Implementation-phase closeout (2026-07-30) — awaiting the matrix

Every implementation slice is COMPLETE and Opus-reviewed; the campaign now waits on the single user gate below.

Register scorecard: goal-enumerated physics stopgaps at ZERO — TS-1, TS-4 (+ its FlatBspQuery twin), TS-5, TS-23, TS-35, TS-46 retired by ports; TS-25 retired on #219 evidence; TS-24→AD-57, TS-40→AD-58 re-argued. AP-7, AP-10, AP-25, AP-71 retired; UN-8 and AD-55 retired by raw-byte proof; AD-25 retired. AP-127 was subsequently retired by #268; the remaining new argued rows are AP-128/129, AD-53/54/55(retired)/56/57/58.

Issues: #72, #153, #167, #255 closed; #116 shape-2 closed /shape-1 narrowed to a probable harness artifact (response layer byte-verified); #165 diagnosed to the render-lag candidate (matrix scenario 8a decides); #166 all four composite deviations landed (scenario 5 decides); #262 apparatus permanently live + 3/3 clean probe logins (scenario 11 decides). Notable finds along the way: ACE's inverted leash-start mapping, ACE's radians/degrees sled-constant bug (cos 10°), the housing-lockout prevalence catch (103,766 restricted cells), and the HouseOwner/Monarch PWD fields that were parsed-and-discarded.

Verification: every slice gated on the complete Release suite; final state 9,977 passed / 0 failed / 4 skipped (the D4 un-skip retired one permanent skip). Suite grew from 8,826 to 9,977 tests over the campaign (+1,151, all conformance/golden/pin coverage).

Final batched connected visual matrix (the ONE user gate)

  1. Burden >100%: run slows, jump shrinks; ~200%: barely moves/jumps.
  2. Repeated jumps drain stamina; low stamina → weak/refused jump; exhaustion behavior matches retail.
  3. Fresh vitae: movement penalty present.
  4. Walk off a cliff/roof edge: slides over like retail, no dead stop.
  5. Downhill jump landing: sled glide + bounce.
  6. Shallow-angle wall graze: lateral slide preserved.
  7. Packed crowd: spacing + shuffle-out unchanged (regression).
  8. Two-client: remote stops at walls without visible penetration; remote ceiling-jump bounces down immediately (#173); Holtburg portal platform step-up (#172); door Use after jumping (#174); closed-door collision matches the visual door (#175); observed-player blips gone (#41).
  9. Locked/barred house: blocked at the threshold.
  10. Wading: slight retail sink-in.
  11. ~20 fresh logins: no run-on-the-spot.
  12. Regression sweep: walk/run/strafe/turn/jump/stairs/doors/water edges feel unchanged from the accepted R6 baseline.

Risk notes

  • P2 and P3 touch the frozen-adjacent transition internals — every subagent prompt must carry the digest's DO-NOT-RETRY table and the no-workarounds rule; 3 failed attempts on any item = stop and build apparatus, per feedback_apparatus_for_physics_bugs.
  • TS-46 changes the collision capsule of every mover; the captured replay fixtures pin behavior — re-baseline only with a recorded retail argument.
  • P1's enchantment-aware skill read is the scope-creep risk; it is bounded to the run/jump query path by this plan.

Live-gate session 2 (2026-07-30) — speed + bounce family landed

The matrix's first live rows surfaced three defects; all three are root-caused, retail-ported, and user-accepted in the same session:

  • #266 CLOSED — run speed: retail MovementSystem::GetRunRate (0x006b0950) treats 800 as an EXACT-EQUALITY sentinel; ACE's >= 800 reading is a misread of the same x87 mush that our P1 port inherited, flat-lining every maxed character at 4.5 (retail-true ~3.70, +21%, vitae-independent). Byte-decoded, fixed at 61e95916; the [stat-chain] live capture proved the vitae→skill chain correct end-to-end. Side-by- side pace vs a retail client accepted by the user.
  • #265/#166 landing-momentum + bounce family — two stacked fixes: (1) c60f6e5d stopped hand-zeroing grounded residual velocity and wired the never-written GroundNormal (roof slides restored); (2) 2d611b2b replaced the AD-25 landing adaptation with the retail mechanism: check_contact (0x0050f5b0) transition seeding, the velocity-free SetPositionInternal commit (0x00515330), and the live 5%-elasticity landing reflect (DEFAULT_ELASTICITY 0.05 @0x007c6a7c). Downhill bounce chain, flat-ground pop, and clean uphill landings all user-accepted ("almost pass with merits"). Investigation + byte-decode record: docs/research/2026-07-30-landing-bounce-family.md.
  • #267 shipped (vitae/buff panel values; attributes vitae-immune). #268 closed 2026-07-31: the complete augmentation chain is shared by panel and Runtime movement; AP-127 is retired. Attributes, secondary attributes, and skills use retail's vitae-excluded green/red comparison. The selected-skill footer now renders per-fragment colors through the shared retained text primitive, using the authored 0x1B palette exactly: #7FFFFF vitae, #00FF00 buff, #FF0000 debuff. TS-8 is also retired: a real live 0x02C2 payload carries its full StatMod through dispatch and changes the effective skill immediately. The user accepted the live colors, values, footer, and immediate row refresh.
  • #269 closed 2026-07-31 — the live 2,184-quantum trace proved the landing reflect and friction math were correct. ACDream omitted retail's OBJECTINFO::kill_velocity before restoring a remembered contact plane in CTransition::validate_transition @ 0x0050AA70, retaining full downhill velocity while repeatedly re-grounding the mover. The exact non-OK-only restore/kill order and final last-known validity overwrite are now ported, focused/full gates pass, and the user accepted repeated slope jumps. Evidence: docs/research/2026-07-31-269-slope-stop-capture.md.
  • #271 closed 2026-07-31 — a bounded stair-side trace proved ACDream could bypass retail's current-position edge back-probe by promoting a stale LastWalkable tread. That made PrecipiceSlide reverse an uphill tangent and rapidly carry the player down the stairs. The two stale-history substitutions are removed; current-walkable, back-probe, and no-walkable outcomes now follow CTransition::edge_slide @ 0x0050B3D0. The exact captured frame is pinned in the existing installed-stair fixture and the complete Release suite passes 10,062 tests / 5 skips. The user accepted repeated uphill runs while pressing into the stair sides. Evidence: docs/research/2026-07-31-271-stair-side-slide-capture.md.

Matrix rows accepted so far: speed parity, roof slide, downhill bounce, flat pop, uphill landing, and #269's slope-stop feel (rows 3/4/5-partial/12-partial). Remaining rows: 1, 2, 6, 7, 8, 9, 10, and 11. The #269 checkpoint passes 4,107 Core tests / 2 skips and 439 Runtime tests / 0 skips; the complete Release suite passes 10,061 tests / 5 skips / 0 failures.