@woosh/meep-engine 3.0.2 → 3.2.0

package/src/engine/physics/RAPIER_REVIEW.md

@@ -1,934 +1,934 @@
-# Deep Technical Review: meep physics vs. Rapier / Parry
-
-A focused, side-by-side comparison of the in-house deterministic JS rigid-body
-engine in `engine/physics/` against [Rapier](https://github.com/dimforge/rapier)
-(Rust dynamics) and its narrowphase library
-[Parry](https://github.com/dimforge/parry). The aim is to surface correctness
-gaps, numerical issues, and structural divergences worth investing in — not to
-recommend a rewrite. Pure-JS, single-threaded and no-SAB are deliberate
-constraints (see `PLAN.md`) and are not listed as gaps.
-
-Repo paths in this document are relative to
-`engine/physics/` for the meep side, and relative to the upstream repo
-(`rapier/src/...`, `parry/src/...`) for the Rapier side.
-
----
-
-## 1. Overall architecture
-
-### 1.1 Pipeline shape
-
-Both engines run the same canonical fixed-step pipeline. The difference is
-mostly granularity and where the boundaries fall:
-
-| Stage | meep `PhysicsSystem.fixedUpdate` | Rapier (per substep inside `VelocitySolver::solve_constraints`) |
-|---|---|---|
-| Integrate velocities (forces + gravity) | semi-implicit Euler, `integrate_velocity.js` | per-substep velocity increment (`solver_vels_increment`) |
-| Refit broadphase (fat AABB) | `compute_fat_world_aabb.js` per awake-body leaf | `BroadPhaseBvh::update`, change-tracked subtree refit |
-| Broadphase pair generation | per-leaf query against static + dynamic BVH, dedup via touched flag | single-tree `BroadPhaseBvh` with pair hashmap |
-| Wake propagation | `__wake_pairs` walks pair list, wakes sleepers | island-manager sleep root traversal |
-| Narrowphase | type-switch dispatcher in `narrowphase_step.js` | dispatch into specialised `contact_manifolds_*` per pair type |
-| Island build | union-find, deterministic CSR layout, separates static/dynamic | DFS active-list propagation, per-island sleep |
-| Solver | per-island PGS+Baumgarte, 10 velocity iters | per-island PGS-S (TGS substepping) + bias/no-bias passes; warm-start |
-| Integrate position | semi-implicit, post-solve | between substeps |
-| Sleep test | per-island atomic, threshold on `\|v\|² + \|ω\|²` | per-island energy threshold, sleep-root traversal |
-| Event diff | `diff_manifolds.js` → Begin/Stay/End | identical model via `ContactEventBuffer` |
-
-The high-level shape is essentially identical. Where meep is closest to
-Rapier is **broadphase refit + pair generation + island building +
-sleep + event dispatch**. Where meep is structurally distinct is **solver
-architecture** (no TGS / split-impulse), **CCD** (speculative margin only),
-and **closed-form narrowphase coverage** (a fraction of Parry's pair
-generators).
-
-### 1.2 Data layout
-
-Both engines use SoA arena storage with generation-tagged stable IDs:
-
-- meep: `BodyStorage.js` — `Int32Array __entities`, `Uint8Array __generations`,
-  packed `(index << 8) | gen`, min-heap free-list (deterministic reuse), dense
-  awake list + reverse map.
-- Rapier: `RigidBodySet` is an `Arena<RigidBody>` (the `Arena` is generic and
-  carries `(index, generation)` handles); `ColliderSet` likewise. The arena's
-  free-slot picking is also LIFO/heap-style for deterministic replay.
-
-**These are structurally the same design.** The differences are:
-
-- meep packs `(index, gen)` into a single u32 with an 8-bit generation; Rapier
-  uses a 64-bit `(u32, u32)` pair. Our wrap-mod-256 generation is fine for
-  games (256 alloc/free cycles before ABA risk on a slot) but is much tighter
-  than Rapier's.
-- meep maintains a **separate dense awake list with reverse map** so the hot
-  loop iterates the active subset directly. Rapier achieves the same via the
-  `IslandManager`'s `awake_islands` list of island ids — different layer of
-  abstraction, same end result.
-- meep stores per-body component pointers in a sparse `__bodies` array
-  (`PhysicsSystem.js:201`) by body index. Rapier returns `RigidBody` refs
-  directly out of the arena. Equivalent.
-
-### 1.3 Broadphase
-
-meep currently runs **two BVHs (static + dynamic)** with fat-AABB refit,
-mirroring Jolt's separation. Rapier was previously `BroadPhaseMultiSap` but
-is now (current master) `BroadPhaseBvh` — a **single unified BVH** with
-change-tracked subtree refits:
-
-```rust
-pub struct BroadPhaseBvh {
-    pub(crate) tree: Bvh,
-    workspace: BvhWorkspace,
-    pairs: HashMap<(ColliderHandle, ColliderHandle), u32>,
-    frame_index: u32,
-    optimization_strategy: BvhOptimizationStrategy,
-}
-```
-
-The directional shift is interesting: **our two-tree split is what Rapier
-moved away from** when it converged on BVH. Rapier's single-tree approach
-relies on `collider.needs_broad_phase_update()` to skip static leaves at
-refit time, which is roughly equivalent to our static-tree never being
-refitted at all. The two trees, however, give us tighter static-side
-queries (the dynamic tree is small even on huge scenes, so the
-dynamic↔static query side of `generate_pairs.js:82-99` walks a balanced
-shallow tree). The trade-off vs. Rapier:
-
-- **Memory:** marginally higher for us — two BVH roots, two free-lists.
-- **Refit cost:** lower for us — we never touch static leaves.
-- **Query cost:** ~same; we pay one extra root-test per query, Rapier pays
-  for static leaf early-outs.
-- **Hot-collider streaming (LOD-load / world streaming):** they're equivalent
-  — both must insert/remove leaves on the active tree.
-
-Our pair de-dup goes through `ManifoldStore.is_touched(slot)` in
-`generate_pairs.js:76,94`. Rapier uses the `HashMap<(handle, handle), u32>`
-directly. Functionally equivalent; ours is allocation-free in steady state
-(open-addressed `PairUint32Map`).
-
-### 1.4 Island structure
-
-Two genuinely different designs:
-
-- **meep** (`IslandBuilder.js`) does **union-find with path halving + union
-  by min-index**, then writes a CSR-laid-out `(body_offsets, body_data,
-  contact_offsets, contact_data)` per frame. Deterministic and re-buildable
-  from scratch each step.
-- **Rapier** (`island_manager/manager.rs`) uses **incremental DFS with
-  sleep-root traversal**. Islands persist across frames; only sleep-state
-  transitions trigger re-traversal, with a `MAX_PER_FRAME_COST` budget to
-  avoid frame spikes.
-
-The Rapier approach is fundamentally laziness-driven: it pays cost only
-when a body changes sleep state. The meep approach pays a small union-find
-cost every frame.
-
-For our workload (many mostly-sleeping bodies + a small active set), our
-rebuild-from-scratch approach is fine because we only iterate awake
-bodies + touched contacts. The union-find pass is `O((|awake_bodies| +
-|touched_contacts|) · α(n))`, dominated by the awake count which is small
-by design. **The deterministic CSR output is a meep-side win** — Rapier
-sacrifices output ordering predictability for incrementality, which makes
-it harder to do bit-exact replays of solver behaviour across two
-instances.
-
-### 1.5 Solver architecture
-
-This is the largest structural divergence.
-
-- **meep**: PGS in `solver/solve_contacts.js`, single velocity-iteration loop,
-  warm-start, Baumgarte position-correction folded into the velocity solve
-  via `bias_n` (`solve_contacts.js:444-449`), Coulomb friction with disk
-  clamp.
-- **Rapier**: PGS-S (TGS substepping) in `velocity_solver.rs`. Two-phase
-  per substep:
-  1. `num_internal_pgs_iterations` of `contact_constraints.solve(...)` —
-     **with bias** (`rhs` includes restitution + position-correction).
-  2. `num_internal_stabilization_iterations` of
-     `contact_constraints.solve_wo_bias(...)` — pure non-penetration,
-     stabilising the bias-pumped velocity.
-  Between substeps Rapier calls `integrate_positions(...)`. The constraint
-  element (`contact_constraint_element.rs`) carries **two impulse fields:
-  `impulse` (current substep) and `impulse_accumulator` (sum across substeps)**
-  and **two RHS fields: `rhs` and `rhs_wo_bias`**. That's the split-impulse
-  + TGS architecture that PLAN.md notes we tried and rolled back.
-
-The relevant Rapier snippet (paraphrased from `contact_constraint_element.rs`):
-
-```rust
-// solve() at iteration k:
-let dvel = dir1·v1 + ... + self.rhs;
-let new_impulse = cfm * (self.impulse - r * dvel).max(0);
-self.impulse = new_impulse;
-// apply lambda = new_impulse - prev_impulse
-```
-
-```rust
-// solve_wo_bias() — same arithmetic but with self.rhs replaced by self.rhs_wo_bias.
-// Position-bias contribution is excluded so the iterates only enforce vn ≥ 0.
-```
-
-The crucial property is that **restitution is encoded once into `rhs` at
-constraint build time and `rhs_wo_bias` does not include it**, so the
-post-bias stabilisation pass cannot subtract a previously-applied
-restitution impulse. That's the fix for PLAN.md's "restitution × warm-start
-clamp" failure mode.
-
-### 1.6 Sleep system
-
-Both have **atomic per-island sleep**. The mechanics differ:
-
-- **meep** (`PhysicsSystem.__sleep_test` at `PhysicsSystem.js:942`): walks
-  islands once per step; if max `|v|² + |ω|²` is below threshold for the
-  whole island and the lowest-stabilising member's sleep_timer crosses the
-  time threshold, threads members into a circular doubly-linked list
-  (`sleep_group_next/prev`), zeroes velocities, and removes from the awake
-  set. Atomic wake walks the chain.
-- **Rapier**: per-body kinetic energy (`activation.update_energy()`) drives
-  a sleep_root candidate flag; the IslandManager extracts sleeping islands
-  by graph traversal from those roots, time-sliced via `MAX_PER_FRAME_COST`.
-  Joint and contact connections re-awaken bodies.
-
-Functionally similar — both wake all members in one frame when any member
-is disturbed. The meep approach is simpler and runs every frame; Rapier's
-is lazier and budget-aware.
-
-### 1.7 Threading
-
-Rapier offers per-island parallel solve via Rayon when the `parallel`
-feature is enabled. meep is deliberately single-threaded (no SAB, no
-workers). The island CSR layout in meep would support a worker-based
-solver but isn't wired up. **Not a gap** — explicit out-of-scope decision.
-
----
-
-## 2. Specific algorithms and trade-offs
-
-### 2.1 Broadphase
-
-| Aspect | meep | Rapier |
-|---|---|---|
-| Structure | Two BVHs (static + dynamic), Jolt-style | One BVH (`BroadPhaseBvh`), change-tracked |
-| Refit | Per awake-body fat AABB (`compute_fat_world_aabb.js:34`) | `collider.needs_broad_phase_update()` then subtree refit |
-| Pair output | `PairList` canonical `(min,max)` | `HashMap<(handle,handle), u32>` |
-| Dedup | `ManifoldStore.is_touched` (one slot per pair) | hashmap key uniqueness + per-frame frame_index |
-| Velocity-pad | `FAT_LINEAR + FAT_VELOCITY_MULTIPLIER * speed * dt` (2 s look-ahead) | configurable; default similar (multiple of step) |
-
-Rapier's optimisation strategy (`SubtreeOptimizer`) does incremental subtree
-re-balancing each frame to amortise the cost of trees becoming
-unbalanced as bodies move. We don't have an equivalent — our BVH's quality
-degrades slowly under sustained motion patterns. Worth investigating for
-the awake set, less so for the static tree which we never touch.
-
-### 2.2 Narrowphase dispatch and contact generators
-
-Parry's narrowphase is a **shape-type-pair dispatch table** with about
-**18 contact-manifold generators**, each a closed-form algorithm for a
-specific shape pair. meep's dispatcher in `narrowphase_step.js:192-353`
-covers **five closed-form pairs** plus GJK+EPA fallback for everything else.
-
-Side-by-side coverage:
-
-| Pair | meep | Parry equivalent |
-|---|---|---|
-| sphere-sphere | ✅ `sphere_sphere_contact.js` | `contact_manifolds_ball_ball.rs` |
-| sphere-box (cuboid) | ✅ `sphere_box_contact.js` | `contact_manifolds_convex_ball.rs` (general) |
-| capsule-sphere | ✅ in `capsule_contacts.js` | `contact_manifolds_convex_ball.rs` |
-| capsule-capsule | ✅ in `capsule_contacts.js` | `contact_manifolds_capsule_capsule.rs` |
-| capsule-box | ✅ multi-contact in `capsule_contacts.js` | `contact_manifolds_cuboid_capsule.rs` |
-| box-box | ✅ SAT + clipping `box_box_manifold.js` | `contact_manifolds_cuboid_cuboid.rs` |
-| sphere-triangle (heightmap/mesh) | ❌ GJK+EPA per triangle | `contact_manifolds_trimesh_shape.rs` dispatches to closed-form |
-| box-triangle | ❌ GJK+EPA per triangle | ✅ `contact_manifolds_cuboid_triangle.rs` |
-| capsule-triangle | ❌ GJK+EPA per triangle | dispatched via convex-vs-trimesh; closed-form exists |
-| half-space (plane) vs convex | ❌ (no plane shape) | ✅ `contact_manifolds_halfspace_pfm.rs` |
-| heightfield | per-triangle GJK+EPA via `heightmap_enumerate_triangles` | `contact_manifolds_heightfield_shape.rs` |
-| convex-convex generic | GJK+EPA | `contact_manifolds_pfm_pfm.rs` (polyhedral-feature-mapping) |
-
-**The big gap is closed-form primitive-vs-triangle.** `PLAN.md` is correct
-that this is the biggest accuracy gap, and Parry's `contact_manifolds_cuboid_triangle.rs`
-is the direct port target. Its algorithm is:
-
-1. SAT over cuboid faces (3 axes), triangle plane (1 axis), and cuboid×edge
-   edge crosses (9 axes in 3D).
-2. Build `PolygonalFeature` from the cuboid's support face and the triangle.
-3. Clip features via `PolygonalFeature::contacts(...)` to produce up to 4
-   contacts.
-
-Crucially Parry's algorithm assigns **feature IDs to each clip output** so
-contact persistence across frames is robust. Our per-triangle GJK+EPA has
-no feature-id story at all (see §2.5).
-
-### 2.3 GJK and EPA
-
-**meep GJK** (`gjk/gjk.js`): Kevin Moran reference implementation, no
-Voronoi simplex — direct case-analysis of triangle / tetrahedron
-sub-regions; 64-iter cap; **no frame-coherent simplex caching**. The
-recently-fixed simplex-shift bugs around `gjk.js:214-264` (clear comments
-in source) suggest this code was carrying subtle correctness issues until
-recently.
-
-**Parry GJK** (`parry/src/query/gjk/gjk.rs`): **Voronoi simplex**
-(`VoronoiSimplex`) — explicit projected-origin-and-reduce step. Frame
-coherence via simplex passed in by caller (so callers can persist it
-between frames; Parry persists in `ContactManifoldsWorkspace`).
-
-The Voronoi simplex is the technically-correct GJK formulation — every
-simplex update projects the origin onto the simplex once and reduces
-based on Voronoi regions, rather than testing each potential reduction
-edge/face manually. The trade-off in our case is small for static-shape
-pairs (boxes don't have many support directions), but meaningful for
-mesh / smooth shapes where the reduction path matters.
-
-**meep EPA** (`gjk/expanding_polytope_algorithm.js`): up to 64 faces, 32
-loose edges, 64 iterations; degeneration is silently swallowed and the
-best-so-far face is returned. The `debugger;` statement at line 136 is a
-leftover that should be removed.
-
-**Parry EPA** (`parry/src/query/epa/epa3.rs`): adaptive `eps_tol` scaled
-by `max(scale_of_coordinates, 1.0)` to prevent spurious failures on
-large-coordinate geometry (fixes parry issue #415). Hard 100-iter cap.
-Both stagnation- and progress-based termination criteria.
-
-The **scale-adaptive tolerance** is a concrete, copyable improvement. Our
-`EPA_TOLERANCE = 0.0001` (line 10) is an absolute threshold — on a
-character-scale shape it's fine but on a kilometre-scale heightmap
-collision it's well below the practical noise floor and may cause early
-"convergence" on a degenerate face.
-
-### 2.4 MPR
-
-Both engines have MPR (`gjk/mpr.js` vs `parry/src/query/contact/`/various).
-Parry uses MPR primarily for **support-map vs support-map proximity tests**
-where GJK is overkill. Our `mpr.js` is used only by `shape_cast.js:364` for
-the normal-recovery probe — it's **not wired into `narrowphase_step.js`** as
-a fallback for non-convergent EPA on smooth shapes (PLAN.md notes this as
-its intended role). That's a low-risk swap: per-pair config to prefer MPR
-for shapes whose support function is smooth (sphere, capsule, possibly
-the unwired cylinder), or as a fallback when EPA returns a non-positive
-depth.
-
-### 2.5 Manifold caching and contact persistence
-
-This is where meep and Parry diverge most subtly.
-
-- **meep** (`contact/ManifoldStore.js`):
-  - Slot keyed by `(idA, idB)` via `PairUint32Map`.
-  - Per-slot meta: count, touched bit, prev_touched bit, grace counter.
-  - Each step `acquire(idA, idB)` flips touched=1. `advance_frame` rolls
-    touched → prev_touched.
-  - Contacts overwritten each frame (`clear_contacts(slot)` at
-    `narrowphase_step.js:687`); only `j_n`, `j_t1`, `j_t2` (warm-start
-    impulses) survive into the next frame.
-  - **There is no per-contact feature-id matching.** Warm-start impulses
-    are written by index `(slot, k)` — if the narrowphase emits contacts
-    in a different order this frame vs. last frame (which can absolutely
-    happen for a rotating box-on-box), the warm-start impulses are applied
-    to the **wrong contact points**.
-
-- **Parry** (`contact_manifold.rs`):
-  ```rust
-  pub struct TrackedContact<Data> {
-      pub local_p1: Vector,
-      pub local_p2: Vector,
-      pub dist: Real,
-      pub fid1: PackedFeatureId,  // feature id on shape 1
-      pub fid2: PackedFeatureId,  // feature id on shape 2
-      pub data: Data,             // user data (warm-start impulses)
-  }
-  ```
-  Each per-pair contact generator (cuboid-cuboid, cuboid-triangle, etc.)
-  produces contacts with **stable feature IDs**: e.g. `face(2) × face(5)`,
-  `edge(0,1) × vertex(3)`, etc. `match_contacts()` then transfers the
-  user data (impulses) onto the corresponding contact in the new
-  manifold by feature-ID equality.
-
-**This is a real correctness bug in our solver.** The warm-start cache
-arithmetic is fine in isolation, but in the rotating-box-on-box case the
-reduction pass in `box_box_manifold.js:498-530` and the candidate
-reduction in `narrowphase_step.js:149-178` produce contact orderings that
-can re-shuffle between frames. The solver then warm-starts contact[0]
-with impulse from a different physical contact, smearing across the
-manifold and producing the kind of low-amplitude jitter that's commonly
-attributed to "PGS being lossy" but is actually warm-start mis-aim.
-
-There are two ways to fix:
-
-1. **Add feature-id columns to the manifold stride**, populate from
-   `box_box_manifold` (the reference + incident face indices + Sutherland-
-   Hodgman edge index already define a feature ID), and match by
-   feature ID when assigning impulses. Parry's `match_contacts` is the
-   reference. Stride goes from 13 → 15 floats. Needs cooperation from
-   each closed-form pair generator.
-2. **Fallback: match by spatial proximity** like Parry's
-   `match_contacts_using_positions`. Cheaper to implement (just a 4×4
-   distance matrix at solver pre-step) but loses precision when contacts
-   are within the matching threshold of each other.
-
-This is the single most impactful **correctness** fix in this review.
-
-### 2.6 Solver: PGS vs PGS-S, Baumgarte vs split-impulse
-
-Rapier carries **two impulses and two RHS values** per constraint element:
-
-```rust
-pub struct ContactConstraintNormalPart<N> {
-    torque_dir1, torque_dir2, ii_torque_dir1, ii_torque_dir2,
-    rhs,             // restitution + position bias
-    rhs_wo_bias,     // pure non-penetration
-    impulse,         // this substep's accumulator
-    impulse_accumulator,  // cumulative across substeps
-    r, r_mat_elts,
-}
-```
-
-Inside a substep:
-
-```
-for _ in 0..num_internal_pgs_iterations:
-    contact_constraints.solve(...)        // uses rhs
-for _ in 0..num_internal_stabilization_iterations:
-    contact_constraints.solve_wo_bias(...) // uses rhs_wo_bias
-```
-
-The two-phase scheme is exactly what defuses the failure modes PLAN.md
-documents (restitution × warm-start, Baumgarte K× amplification):
-
-- Restitution is part of `rhs`, applied once during the bias phase. The
-  no-bias stabilisation pass operates on `rhs_wo_bias` so the
-  restitution impulse cannot be subtracted away by a later iteration
-  seeing "separating velocity".
-- Position correction is part of `rhs`, applied during the bias phase, and
-  the no-bias pass cleans up the post-bias velocity contamination.
-- Forces are integrated once per substep into `solver_vels_increment`
-  outside the iteration loop, so the accumulator-timing issue doesn't
-  arise (each substep gets the right share of the force).
-
-Our `solver/solve_contacts.js:531` does:
-
-```js
-const lambda_n = -m_eff_n * (vn + bias_n);
-const sum_n = j_n_accum + lambda_n;
-const new_j_n = sum_n > 0 ? sum_n : 0;  // CLAMP
-```
-
-with `bias_n` blending Baumgarte position correction and restitution
-(`solve_contacts.js:442-457`). On the very next iteration the same
-contact sees a different `vn`, recomputes the full `(vn + bias_n)` target,
-and can subtract whatever the previous iteration added. Restitution is
-particularly fragile because it's a one-shot impulse that should be
-applied during the contact's *first* solver pass and never adjusted
-again, but our scheme keeps re-targeting it.
-
-**This is the architectural lock that prevents TGS**, as `PLAN.md` documents.
-The minimum fix is to **split `rhs` into `rhs_pos` and `rhs_restitution` +
-`rhs_velocity_only`**, apply the position+restitution bias only on a
-single dedicated iteration (or set of iterations), and then run the rest
-of the solver on the velocity-only target. Even without TGS substepping,
-this would defuse the warm-start × restitution interaction documented in
-`solve_contacts.js:452-456`.
-
-### 2.7 CCD
-
-Rapier has full CCD (`ccd/ccd_solver.rs`):
-
-```rust
-pub struct CCDSolver;   // stateless
-```
-
-with three phases:
-1. `update_ccd_active_flags()` based on velocity threshold.
-2. `predict_impacts_at_next_positions()` — sweeps a body's collider
-   against the broadphase, finds first TOI.
-3. `clamp_motions()` — clamps motion to stop just before impact.
-
-We have speculative-margin-via-fat-AABB only. PLAN.md flags this as a
-backlog item; the falling-tower bench is the concrete failure case (180
-of 1000 bodies tunnel through a 1 cm floor on a 1 km drop).
-
-The shape-cast machinery already in `queries/shape_cast.js` provides the
-TOI primitive. The missing piece is the per-body opt-in CCD path that
-runs after broadphase but before solver: identify high-speed bodies
-(`|v| · dt > min_extent`), sweep their colliders, clamp dt accordingly.
-This is conceptually parallel to `__wake_pairs` — a stage between
-broadphase and narrowphase.
-
-### 2.8 Concave / triangle mesh
-
-Both engines decompose a non-convex shape into triangles overlapping the
-convex's AABB, then run a per-triangle algorithm:
-
-- **meep** (`narrowphase_step.js:368-577`): generic per-triangle GJK+EPA
-  with a one-sided face-normal rejection. The `Triangle3D` support
-  function is degenerate along the face-normal axis (all 3 verts project
-  to the same value), so EPA can't tighten near vertical contacts — the
-  exact failure documented in PLAN.md's known limitations.
-
-- **Parry** (`contact_manifolds_trimesh_shape.rs`): dispatches the
-  per-triangle contact to the appropriate **closed-form** generator
-  (`contact_manifolds_cuboid_triangle.rs`, `contact_manifolds_convex_ball.rs`,
-  etc.). Triangle is treated as a polygonal feature with explicit face,
-  3 edges, 3 vertices and **stable feature IDs** per sub-feature.
-
-The actionable port is Parry's `contact_manifolds_cuboid_triangle.rs` —
-SAT + face clipping — adapted to our `Triangle3D` flyweight and
-emitting into our `(world_a, world_b, normal, depth)` stride. A
-ball-vs-triangle closed form (Voronoi-region-based) is even simpler:
-clamp the sphere centre to the triangle and dispatch on which feature
-(face / edge / vertex) the clamp landed on.
-
-### 2.9 Queries (raycast, shape-cast, overlap)
-
-Parry's `QueryPipeline` mirrors what we have in `queries/`:
-
-| meep | Parry equivalent | Notes |
-|---|---|---|
-| `raycast.js` | `query/ray/...` | meep is broadphase AABB-hit only; PLAN.md flags the narrowphase refinement as TODO |
-| `shape_cast.js` | `shape_cast_support_map_support_map.rs` | meep uses **bisection on GJK overlap**; Parry uses **GJK directional_distance / minkowski_ray_cast** — conservative advancement |
-| `overlap_shape.js` | `query/intersection_test/...` | both use broadphase + GJK |
-
-**The shape_cast approach is genuinely different.** Our bisection is
-robust and works, but is O(`log₂(tMax / tol)`) GJK calls per candidate.
-Parry's GJK directional_distance is **continuous refinement** — it
-advances the ray origin in the Minkowski difference each iteration and
-tightens both bounds simultaneously. Typically converges in 5-15
-iterations vs. our 32-step coarse linear search + up to 32 bisection
-steps. This is a clean port target since we already have all the
-support-function machinery.
-
-### 2.10 Determinism
-
-- **Rapier** advertises cross-platform determinism with f32 by avoiding
-  transcendentals and relying on IEEE-754 round-to-nearest behaviour on
-  the supported platforms. Their integration uses `mul_add`-free math
-  where possible.
-- **meep** is **same-runtime bit-exact** because `Math.sin/cos/exp/log` are
-  ULP-correct but not bit-exact across V8 / SpiderMonkey / JSC. PLAN.md
-  is explicit about this. The codebase's typed-array discipline (direct
-  index writes, no `Vector3.set` on hot paths, min-heap free list, sorted
-  awake-list iteration, canonical pair ordering) is the right
-  infrastructure.
-
-The action item to close the cross-runtime gap is documented in PLAN.md
-(soft-float library for transcendentals). Not a gap in scope of this
-review.
-
----
-
-## 3. In-depth correctness comparison on six key touchpoints
-
-### 3.1 Warm-start impulse correctness across reordered contacts
-
-**meep** `solver/solve_contacts.js:467-476`:
-
-```js
-const j_n  = data[off + 10];
-const j_t1 = data[off + 11];
-const j_t2 = data[off + 12];
-const Px = nx * j_n + t1x * j_t1 + t2x * j_t2;
-// ... apply ...
-```
-
-The warm-start impulses live at fixed slot offsets `(slot, k)` in the
-manifold cache. The new frame's narrowphase calls `manifolds.clear_contacts(slot)`
-(at `narrowphase_step.js:687`), which **zeroes the data array including
-`j_n`, `j_t1`, `j_t2`** (see `ManifoldStore.js:233-239`):
-
-```js
-clear_contacts(slot) {
-    const meta_off = slot * SLOT_META_STRIDE;
-    this.__meta[meta_off + 2] = this.__meta[meta_off + 2] & ~COUNT_MASK;
-    const data_off = slot * SLOT_DATA_STRIDE;
-    this.__data.fill(0, data_off, data_off + SLOT_DATA_STRIDE);
-}
-```
-
-**Bug:** the docstring at `ManifoldStore.js:230` claims "narrowphase
-determines the pair is separated this frame" but `narrowphase_step.js:687`
-calls it **unconditionally** at the start of every narrowphase pass for
-every pair, before the new contacts are computed. So warm-start
-impulses are actively **destroyed on every frame**, every contact, every
-pair. That's exactly the opposite of what warm-starting requires.
-
-Walk-through:
-- Frame N: pair has contact[0] with `j_n = 5.2`. End of frame: solver has
-  applied impulses, data[10] = 5.2.
-- Frame N+1: `acquire()` is a no-op (slot exists, just sets touched).
-- Frame N+1 narrowphase: `clear_contacts(slot)` runs — **5.2 is wiped to 0**.
-- Frame N+1 contacts written via `set_contact`, which **does not touch
-  data[10..12]** (per its own docstring at `ManifoldStore.js:270` "warm-start
-  impulses are preserved") — but they were already zeroed by `clear_contacts`.
-- Solver pre-step at `solve_contacts.js:468-470` reads `j_n=0, j_t1=0, j_t2=0` →
-  no warm-start impulse applied.
-
-**The warm-start cache is effectively cold every frame.** This is a real
-bug — the comment at `ManifoldStore.js:269` "warm-start impulses are
-preserved across calls" describes intended behaviour that the
-`clear_contacts → set_contact` sequence in narrowphase actively breaks.
-
-Even if `clear_contacts` were fixed to preserve `j_n/j_t1/j_t2`, the
-**ordering instability** from §2.5 is the deeper issue: contacts[0]
-this frame is not contact[0] last frame for any non-trivial pair. Need
-feature-ID or position matching.
-
-**Improvement opportunity:**
-1. Make `clear_contacts` either skip the j_n/j_t1/j_t2 stride elements, or
-   accept a flag for "narrowphase contact replacement" vs. "pair separated".
-2. Implement feature-id or position matching at the contact-write site
-   (`narrowphase_step.js:691-700`) — when writing the new contact, find
-   the most-similar previous contact in this slot and **inherit its
-   warm-start impulses**.
-
-### 3.2 Box-box manifold reduction non-determinism vs. Parry's clipping
-
-**meep** `box_box_manifold.js:486-531`: reduce candidates to 4 by
-"deepest, then max-distance-to-kept-set" — a perimeter-maximisation
-heuristic. The ordering depends on numerical tiebreaks in
-`d2 < min_d2` comparisons. For symmetric inputs (two axis-aligned
-unit cubes at integer positions), this can produce different orderings
-in two otherwise-identical PhysicsSystem instances if floating-point
-arithmetic happens to differ by 1 ULP somewhere upstream.
-
-**Parry** `contact_manifolds_cuboid_cuboid.rs`: clipping produces
-**at most 4** candidates by construction (face × face Sutherland-Hodgman
-on a rectangle gives ≤ 8 corners; the four "deepest" are kept by a
-**feature-ID-based selection**, not a metric). Stable orderings.
-
-Our reduction also doesn't guarantee that the **deepest 4** are kept —
-it keeps deepest #1, then maximises spread. A cluster of three deep
-contacts plus one shallow but spread-out contact will lose one of the
-deep ones. For a flat box-on-floor with one corner slightly higher,
-that's the right behaviour. For a tilted box where the corner contacts
-matter, it can drop important contacts.
-
-**Improvement opportunity:** mirror Parry's feature-ID-based 4-of-8 pick
-(each corner of the clipped polygon carries the source edge and incident
-corner id; pick the four corners that span the largest convex hull).
-
-### 3.3 EPA on smooth supports — confirmed degeneration
-
-**meep** `expanding_polytope_algorithm.js:255-263`:
-
-```js
-if (num_loose_edges >= EPA_MAX_NUM_LOOSE_EDGES) {
-    // Polytope degenerated ...
-    // Bail out and return the closest-face result accumulated so far ...
-    break;
-}
-```
-
-The fallback path returns the **last closest face's normal × `dot_p_search_dir`**
-as the depth — but `dot_p_search_dir` was computed from a new support point
-that the polytope-update loop just **discarded** (because it can't fit it
-into the polytope without exceeding the cap). That's a mismatch: the depth
-is the projection of a vertex that isn't on the polytope, while the
-direction is the polytope's closest-face normal. The two should be consistent.
-
-**Parry** `epa3.rs`: scale-adaptive tolerance + 100-iteration cap; when
-it bails, it returns the closest face's barycentric-projected point, which
-is by construction a Minkowski-difference point. Output is a true MTV.
-
-**Improvement opportunity:**
-1. On the EPA bail path, return `closest_face_normal * min_dist` (the
-   stored `min_dist` for the chosen closest face), not
-   `normal * dot_p_search_dir`. `min_dist` is the actual signed distance
-   from origin to the closest face plane, which is the right MTV magnitude.
-2. Scale `EPA_TOLERANCE` by `max(scale_of_coordinates, 1.0)` — same fix
-   Parry applied for issue #415. The coordinate scale is the max of the
-   incoming simplex's coordinate magnitudes.
-3. Remove the leftover `debugger;` at `expanding_polytope_algorithm.js:136`.
-
-### 3.4 Concave dispatch: face-normal rejection + concave-vs-concave
-
-**meep** `narrowphase_step.js:519`:
-
-```js
-if (ex * fnx_w + ey * fny_w + ez * fnz_w <= 0) continue;
-```
-
-This is the "convex on outward side of triangle" test. **Correctness
-concern:** for a sphere centred slightly above a flat heightmap, the
-sphere-vs-triangle GJK+EPA may report an MTV that's near-perpendicular to
-the face (good — `fnx_w · ex > 0`), but for cases where the sphere is
-overlapping two adjacent triangles (one slanted, one horizontal), the
-sphere-vs-slanted-triangle's EPA can return an MTV that points *into*
-the slanted face. The rejection then skips that triangle, leaving only
-the horizontal-face contact — which is correct here. But:
-
-The face-normal rejection runs **after** sign-canonicalisation against the
-triangle centroid → convex centre vector. For a thin shape (capsule
-oriented along the slope), the centroid-to-centre vector might be near-
-parallel to the face, and the sign canonicalisation flips inconsistently
-between adjacent triangles, producing alternating contact directions
-across the heightmap. The `compute_penetration.js` half-space pre-test
-sidesteps this; the narrowphase doesn't.
-
-**Improvement opportunity:** for primitive-vs-triangle, use the
-half-space approach (project sphere centre / box corner / capsule
-endpoints onto the triangle plane) like `compute_penetration.js` does.
-This is roughly half the work of a closed-form ball-vs-triangle and
-already validated in our codebase. The full closed-form (§2.8) is
-strictly better — see backlog.
-
-### 3.5 GJK iteration tolerance and convergence detection
-
-**meep** `gjk.js:96-128`: 64-iter cap, **no progress check between
-iterations**. The only termination is "origin enclosed in tetrahedron"
-(return true) or "new support doesn't pass the origin" (return false).
-Non-progressing inputs (degenerate Minkowski difference at exactly
-touching, smooth-on-smooth tangent contact) run to 64 iterations and
-return false.
-
-**Parry** `gjk.rs`: progress check — if the new support's signed distance
-to the previous closest-feature plane improves by less than `eps`, declare
-convergence. Faster termination on near-touching pairs (sphere tangent to
-sphere); same false return on actual non-overlap but with fewer
-iterations.
-
-**Improvement opportunity:** add a progress check on the new support
-point's projection along the search direction. If `dot(new_support, dir)`
-is within tolerance of the previous iteration's value, declare convergence.
-On the practical level this affects performance (smooth-on-smooth) more
-than correctness (we already return false on the iteration cap), but
-falling out earlier on tangent contact is a real perf+stability win for
-the broadphase-edge case.
-
-### 3.6 Shape-cast bisection vs. GJK conservative advancement
-
-**meep** `queries/shape_cast.js:267-297`: AABB slab interval narrowing +
-32-step coarse linear search + up to 32 bisection steps + MPR normal
-recovery.
-
-**Parry** `shape_cast_support_map_support_map.rs` → `gjk::directional_distance`:
-**continuous refinement** in Minkowski space, simultaneously tightening
-both bounds. Converges in ~5-15 iterations on typical inputs because each
-iteration uses geometric structure (the support hyperplane at the current
-ray position) rather than blind bisection.
-
-Trade-offs:
-- Bisection is **simpler and dead-reliable** — no degenerate cases.
-- Conservative advancement is **substantially fewer GJK calls** —
-  Parry's typical TOI query is in the 10s of microseconds; ours is
-  64-step bisection per candidate.
-
-For our workload (kinematic character controllers — a few queries per
-frame), the perf delta is small. **Not a high-priority change** but
-worth knowing about. The normal recovery via MPR probe-step is fine
-(`shape_cast.js:350-381`); Parry's TOI returns the contact normal as
-part of the same conservative-advancement loop, so there's no separate
-probe needed.
-
-### 3.7 Friction: tangent disk vs. Coulomb pyramid
-
-**meep** `solver/friction_cone.js` (called from `solve_contacts.js:569`):
-two-tangent disk clamp — limits `sqrt(t1² + t2²) ≤ μ·j_n`. Coulomb cone
-in the velocity-impulse plane.
-
-**Parry**: looks similar in the `ContactConstraintTangentPart::solve` —
-Vector2 impulse clamped against `limit = mu * normal_part.impulse`.
-Functionally the same.
-
-No actionable difference; mark as parity.
-
-### 3.8 Body-centre application point fallback in GJK+EPA path
-
-**meep** `narrowphase_step.js:640-644`:
-
-```js
-return append_contact(count,
-    trA.position.x, trA.position.y, trA.position.z,
-    trB.position.x, trB.position.y, trB.position.z,
-    nx, ny, nz,
-    depth);
-```
-
-The fallback contact application points are **body centres**, with the
-detailed explanation in the inline comment (`narrowphase_step.js:623-639`)
-about flat-face support being multi-valued. This is correct in spirit but
-not in detail: for a tilted box on a flat floor with the GJK+EPA path
-(e.g. a non-cube convex hull on a heightmap triangle), the contact
-"point" sitting at the floor's centre rather than under the box produces
-the wrong torque arm. Specifically `r × n` is large and in the wrong
-direction. The Baumgarte position correction gets dampened by the large
-effective mass denominator and the box separates slowly.
-
-**Parry** computes contact points as **support-function witnesses**: for
-each shape, the support point in the contact normal direction is the
-contact application point. For flat-face shapes the witness is one of
-the (possibly-multi-valued) corners, and Parry deals with the
-multi-valued case by **picking a face when the support is multi-valued
-in the support direction**. Their `support_face` returns a
-`PolygonalFeature` rather than a point.
-
-**Improvement opportunity:** in the GJK+EPA fallback path, compute the
-contact point as `(posed_a.support(epa_dir) + posed_b.support(-epa_dir)) / 2`
-**modulo a check for flat-face support**: if `|support_a - centre_a|` or
-`|support_b - centre_b|` is suspiciously large relative to the shape's
-characteristic length, fall back to body centres. This narrows the body-
-centre fallback to its actually-correct domain (flat-faced shapes where
-the support is multi-valued) instead of using it for everything.
-
----
-
-## 4. Simplicity and uniformity
-
-### 4.1 What's clean
-
-- **`BodyStorage`** (`body/BodyStorage.js`): SoA, generation packing,
-  deterministic free-list. Every primitive operation is one block of
-  index arithmetic. Hard to make simpler without losing functionality.
-- **`PairUint32Map`**: Robin Hood hashing in a single typed-array. No
-  allocation in steady state.
-- **`union_find.js`**: path halving in 8 lines. Minimal code.
-- **`integrate_velocity.js`**: 80 lines, no surprises, well-commented on
-  the implicit-decay damping recipe.
-- **`compute_fat_world_aabb.js`**: 60 lines, single-purpose, all constants
-  named.
-
-### 4.2 What's specialised in ways that earn their keep
-
-- **The two-BVH split** (`PhysicsSystem.staticBvh` + `dynamicBvh`): cheap
-  cognitively, real perf win for the "mostly-static world" target. Static
-  tree is built once and never touched.
-- **Atomic-island sleep with circular doubly-linked sleep groups**: the
-  alternative ("rebuild from scratch") would be much simpler but loses
-  the property that a hit at the base of a 100-stack wakes the whole
-  stack in one frame.
-- **Per-island CSR layout from `IslandBuilder`**: enables solver,
-  sleep-test, and event-dispatch to all iterate the same data structure.
-  Genuine reuse.
-- **`Triangle3D` flyweight + `triangle_buffer` reuse in the concave
-  dispatch**: zero allocation per triangle. The alternative (allocating
-  triangle instances) would be a hotspot.
-
-### 4.3 What's specialised in ways that don't earn their keep
-
-- **The narrowphase dispatcher** in `narrowphase_step.js:192-353` is a big
-  if-else chain on `isUnitSphereShape3D`, `isBoxShape3D`, etc., with
-  per-pair flipping of A/B identity (`isSphereA ? trA : trB` etc.). It's
-  legible but doesn't scale gracefully — adding cylinder support is
-  another ~50 lines of switch arms. Parry's approach is a **shape-type
-  pair → dispatch function** lookup, which is one indirection but
-  trivially extensible. A small registry table indexed by
-  `(shape_a_type << 8) | shape_b_type` would replace the entire chain
-  with one indirection. (NB: the type-tag identification via
-  `isUnitSphereShape3D === true` works in our JS world but the registry
-  generalises trivially.)
-- **Quaternion conjugate-rotate-conjugate boilerplate** in 5+ files
-  (`sphere_box_contact.js:48-57`, `capsule_contacts.js:39-71`,
-  `narrowphase_step.js:467-475`, `PosedShape3D.js:67-93`). Each
-  reimplements the same `v_world = q · v_local · q*` arithmetic inline.
-  Extracting a single `quaternion_apply` and trusting V8 to inline it
-  would clean the code at no cost. Parry has `Isometry` for exactly this.
-- **EPA's `loose_edges` array** (`expanding_polytope_algorithm.js:41`):
-  the `copyWithin` based edge removal scales poorly past ~32 edges and
-  the wrapping logic is fragile. Parry's adjacency-graph EPA is harder
-  to write but more robust. Worth keeping the current implementation
-  until it bites.
-
-### 4.4 Adding a new shape pair / constraint / query
-
-- **New shape pair (e.g. cylinder-box):** add a closed-form handler in
-  `narrowphase/cylinder_box_contact.js`, add an `isCylinderShape3D` type
-  tag, add a branch in `narrowphase_step.js` dispatch. ~3 files,
-  ~100-200 lines, mechanical. Parry's equivalent is also mechanical but
-  more strongly typed.
-- **New constraint (e.g. distance joint):** would need to thread `joints`
-  through the island builder + solver. PLAN.md notes "the solver loop is
-  already set up to iterate contacts ∪ joints; only constraint pre-step +
-  warm-start hook is missing." This is structurally easier than Parry
-  because we already iterate per-island.
-- **New query (e.g. sphere-cast):** `queries/sphere_cast.js` next to
-  `shape_cast.js` — but `shape_cast` is already generic, so this would
-  be a thin wrapper. Parry has both `cast_ray` and `cast_shape` for the
-  same reason.
-
-### 4.5 Recommended structural simplifications
-
-1. **Extract a `quaternion_apply(out, off, vx, vy, vz, qx, qy, qz, qw)`
-   helper** in `core/geom/3d/quaternion/`. Replace every inline copy in
-   the narrowphase.
-2. **Replace the narrowphase dispatcher's if-else chain with a small
-   typed dispatch table.** Adding shapes becomes a one-line entry.
-3. **Move warm-start impulse persistence out of the
-   `clear_contacts` path** (§3.1) — narrowphase calls a different method
-   `replace_contacts(slot, new_count)` that leaves j_n/j_t1/j_t2 alone,
-   and `clear_contacts` continues to be the "this pair has separated"
-   reset.
-4. **Centralise contact application-point selection** (§3.8) — a single
-   `contact_point_for_pair(shapeA, shapeB, epa_dir)` that knows when to
-   use witnesses vs. body centres, rather than scattering the heuristic
-   across narrowphase paths.
-
----
-
-## 5. Headline findings, ranked
-
-1. **Warm-start cache is effectively cold every frame** (§3.1) due to
-   `clear_contacts(slot)` running unconditionally at the start of every
-   narrowphase pass. Real bug. Cheap fix.
-2. **Contact persistence lacks feature-id matching** (§2.5). When the
-   narrowphase emits manifold contacts in a different order this frame
-   vs. last, warm-start impulses apply to wrong contact points. Parry's
-   `TrackedContact { fid1, fid2, data }` is the model.
-3. **Solver is structurally locked out of TGS by combined Baumgarte+
-   restitution bias in `rhs`** (§2.6). The Rapier `rhs` vs `rhs_wo_bias`
-   split + impulse_accumulator architecture defuses the three failure
-   modes documented in PLAN.md. Significant work but unblocks the
-   biggest performance roadmap item.
-4. **Closed-form primitive-vs-triangle missing** (§2.8). PLAN.md flags as
-   the single biggest accuracy gap. Parry's `contact_manifolds_cuboid_triangle.rs`
-   + a Voronoi-region ball-vs-triangle are the port targets.
-5. **EPA scale-adaptive tolerance and consistent fallback** (§3.3). The
-   `EPA_TOLERANCE = 0.0001` is an absolute constant; should scale by
-   coordinate magnitude (Parry issue #415's fix). On the EPA bail path,
-   return `min_dist * face_normal`, not the inconsistent
-   `dot_p_search_dir * face_normal`. Also drop the leftover `debugger;`.
-6. **Concave-side per-triangle GJK+EPA cannot converge on degenerate
-   `Triangle3D` support** (§3.4). Already documented in PLAN.md. The
-   sphere-vs-triangle closed-form is the smallest first step (Voronoi
-   region clamp) and removes a class of skipped tests.
-7. **MPR is not wired into the narrowphase as an EPA fallback**.
-   PLAN.md correctly identifies this — the swap is in `narrowphase_step.js:580-644`
-   (replace the EPA call with `if (mpr(...)) ... else if (epa(...))` or
-   prefer MPR on smooth-shape pairs). The code already exists.
-8. **Box-box reduction can drop deep contacts in favour of spread**
-   (§3.2). Mirror Parry's feature-ID-based 4-of-8 corner pick on the
-   clipped polygon.
-9. **GJK lacks progress-based convergence detection** (§3.5). Adds a
-   single dot-product check per iteration; speeds up tangent contacts
-   and provides earlier-out on degenerate Minkowski differences.
-10. **CCD is speculative-margin-only**. Already in backlog. Rapier's
-    `predict_impacts_at_next_positions → clamp_motions` two-phase TOI
-    sweep is the architecture; our `shape_cast.js` already has the TOI
-    primitive.
-
----
-
-## Appendix A: file-by-file Rapier/Parry references
-
-| meep file | Rapier/Parry counterpart |
-|---|---|
-| `ecs/PhysicsSystem.js` | `rapier/src/pipeline/physics_pipeline.rs` |
-| `body/BodyStorage.js` | `rapier/src/dynamics/rigid_body_set.rs` |
-| `broadphase/generate_pairs.js` | `rapier/src/geometry/broad_phase_bvh.rs` |
-| `broadphase/compute_fat_world_aabb.js` | (inline in `broad_phase_bvh.rs` refit logic) |
-| `contact/ManifoldStore.js` | `parry/src/query/contact_manifolds/contact_manifold.rs` |
-| `events/diff_manifolds.js` | `rapier/src/pipeline/event_handler.rs` |
-| `island/IslandBuilder.js` | `rapier/src/dynamics/island_manager/manager.rs` |
-| `core/collection/union-find/union_find.js` | (Rapier uses DFS, not union-find) |
-| `gjk/gjk.js` | `parry/src/query/gjk/gjk.rs` (`VoronoiSimplex` based) |
-| `gjk/expanding_polytope_algorithm.js` | `parry/src/query/epa/epa3.rs` |
-| `gjk/mpr.js` | `parry/src/query/mpr/...` |
-| `narrowphase/sphere_sphere_contact.js` | `parry/src/query/contact/contact_ball_ball.rs` |
-| `narrowphase/sphere_box_contact.js` | `parry/src/query/contact_manifolds/contact_manifolds_convex_ball.rs` |
-| `narrowphase/capsule_contacts.js` | `parry/src/query/contact_manifolds/contact_manifolds_capsule_capsule.rs`, `..._cuboid_capsule.rs` |
-| `narrowphase/box_box_manifold.js` | `parry/src/query/contact_manifolds/contact_manifolds_cuboid_cuboid.rs` |
-| `narrowphase/narrowphase_step.js` (concave path) | `parry/src/query/contact_manifolds/contact_manifolds_trimesh_shape.rs` |
-| `narrowphase/decomposition/heightmap_enumerate_triangles.js` | `parry/src/query/contact_manifolds/contact_manifolds_heightfield_shape.rs` |
-| `narrowphase/compute_penetration.js` | `parry/src/query/contact/contact_shape_shape.rs` |
-| `solver/solve_contacts.js` | `rapier/src/dynamics/solver/velocity_solver.rs` + `rapier/src/dynamics/solver/contact_constraint/contact_constraint_element.rs` |
-| `solver/friction_cone.js` | (inline in `ContactConstraintTangentPart::solve`) |
-| `integration/integrate_velocity.js` | `rapier/src/dynamics/solver/velocity_solver.rs` (between-substep integration) |
-| `integration/integrate_position.js` | `rapier/src/dynamics/solver/velocity_solver.rs` `integrate_positions` |
-| `inertia/world_inverse_inertia.js` | `rapier/src/dynamics/solver/...` (inline in constraint setup) |
-| `queries/raycast.js` | `parry/src/query/ray/...` |
-| `queries/shape_cast.js` | `parry/src/query/shape_cast/shape_cast_support_map_support_map.rs` (uses GJK conservative advancement) |
-| `queries/overlap_shape.js` | `parry/src/query/intersection_test/...` |
+# Deep Technical Review: meep physics vs. Rapier / Parry
+
+A focused, side-by-side comparison of the in-house deterministic JS rigid-body
+engine in `engine/physics/` against [Rapier](https://github.com/dimforge/rapier)
+(Rust dynamics) and its narrowphase library
+[Parry](https://github.com/dimforge/parry). The aim is to surface correctness
+gaps, numerical issues, and structural divergences worth investing in — not to
+recommend a rewrite. Pure-JS, single-threaded and no-SAB are deliberate
+constraints (see `PLAN.md`) and are not listed as gaps.
+
+Repo paths in this document are relative to
+`engine/physics/` for the meep side, and relative to the upstream repo
+(`rapier/src/...`, `parry/src/...`) for the Rapier side.
+
+---
+
+## 1. Overall architecture
+
+### 1.1 Pipeline shape
+
+Both engines run the same canonical fixed-step pipeline. The difference is
+mostly granularity and where the boundaries fall:
+
+| Stage | meep `PhysicsSystem.fixedUpdate` | Rapier (per substep inside `VelocitySolver::solve_constraints`) |
+|---|---|---|
+| Integrate velocities (forces + gravity) | semi-implicit Euler, `integrate_velocity.js` | per-substep velocity increment (`solver_vels_increment`) |
+| Refit broadphase (fat AABB) | `compute_fat_world_aabb.js` per awake-body leaf | `BroadPhaseBvh::update`, change-tracked subtree refit |
+| Broadphase pair generation | per-leaf query against static + dynamic BVH, dedup via touched flag | single-tree `BroadPhaseBvh` with pair hashmap |
+| Wake propagation | `__wake_pairs` walks pair list, wakes sleepers | island-manager sleep root traversal |
+| Narrowphase | type-switch dispatcher in `narrowphase_step.js` | dispatch into specialised `contact_manifolds_*` per pair type |
+| Island build | union-find, deterministic CSR layout, separates static/dynamic | DFS active-list propagation, per-island sleep |
+| Solver | per-island PGS+Baumgarte, 10 velocity iters | per-island PGS-S (TGS substepping) + bias/no-bias passes; warm-start |
+| Integrate position | semi-implicit, post-solve | between substeps |
+| Sleep test | per-island atomic, threshold on `\|v\|² + \|ω\|²` | per-island energy threshold, sleep-root traversal |
+| Event diff | `diff_manifolds.js` → Begin/Stay/End | identical model via `ContactEventBuffer` |
+
+The high-level shape is essentially identical. Where meep is closest to
+Rapier is **broadphase refit + pair generation + island building +
+sleep + event dispatch**. Where meep is structurally distinct is **solver
+architecture** (no TGS / split-impulse), **CCD** (speculative margin only),
+and **closed-form narrowphase coverage** (a fraction of Parry's pair
+generators).
+
+### 1.2 Data layout
+
+Both engines use SoA arena storage with generation-tagged stable IDs:
+
+- meep: `BodyStorage.js` — `Int32Array __entities`, `Uint8Array __generations`,
+  packed `(index << 8) | gen`, min-heap free-list (deterministic reuse), dense
+  awake list + reverse map.
+- Rapier: `RigidBodySet` is an `Arena<RigidBody>` (the `Arena` is generic and
+  carries `(index, generation)` handles); `ColliderSet` likewise. The arena's
+  free-slot picking is also LIFO/heap-style for deterministic replay.
+
+**These are structurally the same design.** The differences are:
+
+- meep packs `(index, gen)` into a single u32 with an 8-bit generation; Rapier
+  uses a 64-bit `(u32, u32)` pair. Our wrap-mod-256 generation is fine for
+  games (256 alloc/free cycles before ABA risk on a slot) but is much tighter
+  than Rapier's.
+- meep maintains a **separate dense awake list with reverse map** so the hot
+  loop iterates the active subset directly. Rapier achieves the same via the
+  `IslandManager`'s `awake_islands` list of island ids — different layer of
+  abstraction, same end result.
+- meep stores per-body component pointers in a sparse `__bodies` array
+  (`PhysicsSystem.js:201`) by body index. Rapier returns `RigidBody` refs
+  directly out of the arena. Equivalent.
+
+### 1.3 Broadphase
+
+meep currently runs **two BVHs (static + dynamic)** with fat-AABB refit,
+mirroring Jolt's separation. Rapier was previously `BroadPhaseMultiSap` but
+is now (current master) `BroadPhaseBvh` — a **single unified BVH** with
+change-tracked subtree refits:
+
+```rust
+pub struct BroadPhaseBvh {
+    pub(crate) tree: Bvh,
+    workspace: BvhWorkspace,
+    pairs: HashMap<(ColliderHandle, ColliderHandle), u32>,
+    frame_index: u32,
+    optimization_strategy: BvhOptimizationStrategy,
+}
+```
+
+The directional shift is interesting: **our two-tree split is what Rapier
+moved away from** when it converged on BVH. Rapier's single-tree approach
+relies on `collider.needs_broad_phase_update()` to skip static leaves at
+refit time, which is roughly equivalent to our static-tree never being
+refitted at all. The two trees, however, give us tighter static-side
+queries (the dynamic tree is small even on huge scenes, so the
+dynamic↔static query side of `generate_pairs.js:82-99` walks a balanced
+shallow tree). The trade-off vs. Rapier:
+
+- **Memory:** marginally higher for us — two BVH roots, two free-lists.
+- **Refit cost:** lower for us — we never touch static leaves.
+- **Query cost:** ~same; we pay one extra root-test per query, Rapier pays
+  for static leaf early-outs.
+- **Hot-collider streaming (LOD-load / world streaming):** they're equivalent
+  — both must insert/remove leaves on the active tree.
+
+Our pair de-dup goes through `ManifoldStore.is_touched(slot)` in
+`generate_pairs.js:76,94`. Rapier uses the `HashMap<(handle, handle), u32>`
+directly. Functionally equivalent; ours is allocation-free in steady state
+(open-addressed `PairUint32Map`).
+
+### 1.4 Island structure
+
+Two genuinely different designs:
+
+- **meep** (`IslandBuilder.js`) does **union-find with path halving + union
+  by min-index**, then writes a CSR-laid-out `(body_offsets, body_data,
+  contact_offsets, contact_data)` per frame. Deterministic and re-buildable
+  from scratch each step.
+- **Rapier** (`island_manager/manager.rs`) uses **incremental DFS with
+  sleep-root traversal**. Islands persist across frames; only sleep-state
+  transitions trigger re-traversal, with a `MAX_PER_FRAME_COST` budget to
+  avoid frame spikes.
+
+The Rapier approach is fundamentally laziness-driven: it pays cost only
+when a body changes sleep state. The meep approach pays a small union-find
+cost every frame.
+
+For our workload (many mostly-sleeping bodies + a small active set), our
+rebuild-from-scratch approach is fine because we only iterate awake
+bodies + touched contacts. The union-find pass is `O((|awake_bodies| +
+|touched_contacts|) · α(n))`, dominated by the awake count which is small
+by design. **The deterministic CSR output is a meep-side win** — Rapier
+sacrifices output ordering predictability for incrementality, which makes
+it harder to do bit-exact replays of solver behaviour across two
+instances.
+
+### 1.5 Solver architecture
+
+This is the largest structural divergence.
+
+- **meep**: PGS in `solver/solve_contacts.js`, single velocity-iteration loop,
+  warm-start, Baumgarte position-correction folded into the velocity solve
+  via `bias_n` (`solve_contacts.js:444-449`), Coulomb friction with disk
+  clamp.
+- **Rapier**: PGS-S (TGS substepping) in `velocity_solver.rs`. Two-phase
+  per substep:
+  1. `num_internal_pgs_iterations` of `contact_constraints.solve(...)` —
+     **with bias** (`rhs` includes restitution + position-correction).
+  2. `num_internal_stabilization_iterations` of
+     `contact_constraints.solve_wo_bias(...)` — pure non-penetration,
+     stabilising the bias-pumped velocity.
+  Between substeps Rapier calls `integrate_positions(...)`. The constraint
+  element (`contact_constraint_element.rs`) carries **two impulse fields:
+  `impulse` (current substep) and `impulse_accumulator` (sum across substeps)**
+  and **two RHS fields: `rhs` and `rhs_wo_bias`**. That's the split-impulse
+  + TGS architecture that PLAN.md notes we tried and rolled back.
+
+The relevant Rapier snippet (paraphrased from `contact_constraint_element.rs`):
+
+```rust
+// solve() at iteration k:
+let dvel = dir1·v1 + ... + self.rhs;
+let new_impulse = cfm * (self.impulse - r * dvel).max(0);
+self.impulse = new_impulse;
+// apply lambda = new_impulse - prev_impulse
+```
+
+```rust
+// solve_wo_bias() — same arithmetic but with self.rhs replaced by self.rhs_wo_bias.
+// Position-bias contribution is excluded so the iterates only enforce vn ≥ 0.
+```
+
+The crucial property is that **restitution is encoded once into `rhs` at
+constraint build time and `rhs_wo_bias` does not include it**, so the
+post-bias stabilisation pass cannot subtract a previously-applied
+restitution impulse. That's the fix for PLAN.md's "restitution × warm-start
+clamp" failure mode.
+
+### 1.6 Sleep system
+
+Both have **atomic per-island sleep**. The mechanics differ:
+
+- **meep** (`PhysicsSystem.__sleep_test` at `PhysicsSystem.js:942`): walks
+  islands once per step; if max `|v|² + |ω|²` is below threshold for the
+  whole island and the lowest-stabilising member's sleep_timer crosses the
+  time threshold, threads members into a circular doubly-linked list
+  (`sleep_group_next/prev`), zeroes velocities, and removes from the awake
+  set. Atomic wake walks the chain.
+- **Rapier**: per-body kinetic energy (`activation.update_energy()`) drives
+  a sleep_root candidate flag; the IslandManager extracts sleeping islands
+  by graph traversal from those roots, time-sliced via `MAX_PER_FRAME_COST`.
+  Joint and contact connections re-awaken bodies.
+
+Functionally similar — both wake all members in one frame when any member
+is disturbed. The meep approach is simpler and runs every frame; Rapier's
+is lazier and budget-aware.
+
+### 1.7 Threading
+
+Rapier offers per-island parallel solve via Rayon when the `parallel`
+feature is enabled. meep is deliberately single-threaded (no SAB, no
+workers). The island CSR layout in meep would support a worker-based
+solver but isn't wired up. **Not a gap** — explicit out-of-scope decision.
+
+---
+
+## 2. Specific algorithms and trade-offs
+
+### 2.1 Broadphase
+
+| Aspect | meep | Rapier |
+|---|---|---|
+| Structure | Two BVHs (static + dynamic), Jolt-style | One BVH (`BroadPhaseBvh`), change-tracked |
+| Refit | Per awake-body fat AABB (`compute_fat_world_aabb.js:34`) | `collider.needs_broad_phase_update()` then subtree refit |
+| Pair output | `PairList` canonical `(min,max)` | `HashMap<(handle,handle), u32>` |
+| Dedup | `ManifoldStore.is_touched` (one slot per pair) | hashmap key uniqueness + per-frame frame_index |
+| Velocity-pad | `FAT_LINEAR + FAT_VELOCITY_MULTIPLIER * speed * dt` (2 s look-ahead) | configurable; default similar (multiple of step) |
+
+Rapier's optimisation strategy (`SubtreeOptimizer`) does incremental subtree
+re-balancing each frame to amortise the cost of trees becoming
+unbalanced as bodies move. We don't have an equivalent — our BVH's quality
+degrades slowly under sustained motion patterns. Worth investigating for
+the awake set, less so for the static tree which we never touch.
+
+### 2.2 Narrowphase dispatch and contact generators
+
+Parry's narrowphase is a **shape-type-pair dispatch table** with about
+**18 contact-manifold generators**, each a closed-form algorithm for a
+specific shape pair. meep's dispatcher in `narrowphase_step.js:192-353`
+covers **five closed-form pairs** plus GJK+EPA fallback for everything else.
+
+Side-by-side coverage:
+
+| Pair | meep | Parry equivalent |
+|---|---|---|
+| sphere-sphere | ✅ `sphere_sphere_contact.js` | `contact_manifolds_ball_ball.rs` |
+| sphere-box (cuboid) | ✅ `sphere_box_contact.js` | `contact_manifolds_convex_ball.rs` (general) |
+| capsule-sphere | ✅ in `capsule_contacts.js` | `contact_manifolds_convex_ball.rs` |
+| capsule-capsule | ✅ in `capsule_contacts.js` | `contact_manifolds_capsule_capsule.rs` |
+| capsule-box | ✅ multi-contact in `capsule_contacts.js` | `contact_manifolds_cuboid_capsule.rs` |
+| box-box | ✅ SAT + clipping `box_box_manifold.js` | `contact_manifolds_cuboid_cuboid.rs` |
+| sphere-triangle (heightmap/mesh) | ❌ GJK+EPA per triangle | `contact_manifolds_trimesh_shape.rs` dispatches to closed-form |
+| box-triangle | ❌ GJK+EPA per triangle | ✅ `contact_manifolds_cuboid_triangle.rs` |
+| capsule-triangle | ❌ GJK+EPA per triangle | dispatched via convex-vs-trimesh; closed-form exists |
+| half-space (plane) vs convex | ❌ (no plane shape) | ✅ `contact_manifolds_halfspace_pfm.rs` |
+| heightfield | per-triangle GJK+EPA via `heightmap_enumerate_triangles` | `contact_manifolds_heightfield_shape.rs` |
+| convex-convex generic | GJK+EPA | `contact_manifolds_pfm_pfm.rs` (polyhedral-feature-mapping) |
+
+**The big gap is closed-form primitive-vs-triangle.** `PLAN.md` is correct
+that this is the biggest accuracy gap, and Parry's `contact_manifolds_cuboid_triangle.rs`
+is the direct port target. Its algorithm is:
+
+1. SAT over cuboid faces (3 axes), triangle plane (1 axis), and cuboid×edge
+   edge crosses (9 axes in 3D).
+2. Build `PolygonalFeature` from the cuboid's support face and the triangle.
+3. Clip features via `PolygonalFeature::contacts(...)` to produce up to 4
+   contacts.
+
+Crucially Parry's algorithm assigns **feature IDs to each clip output** so
+contact persistence across frames is robust. Our per-triangle GJK+EPA has
+no feature-id story at all (see §2.5).
+
+### 2.3 GJK and EPA
+
+**meep GJK** (`gjk/gjk.js`): Kevin Moran reference implementation, no
+Voronoi simplex — direct case-analysis of triangle / tetrahedron
+sub-regions; 64-iter cap; **no frame-coherent simplex caching**. The
+recently-fixed simplex-shift bugs around `gjk.js:214-264` (clear comments
+in source) suggest this code was carrying subtle correctness issues until
+recently.
+
+**Parry GJK** (`parry/src/query/gjk/gjk.rs`): **Voronoi simplex**
+(`VoronoiSimplex`) — explicit projected-origin-and-reduce step. Frame
+coherence via simplex passed in by caller (so callers can persist it
+between frames; Parry persists in `ContactManifoldsWorkspace`).
+
+The Voronoi simplex is the technically-correct GJK formulation — every
+simplex update projects the origin onto the simplex once and reduces
+based on Voronoi regions, rather than testing each potential reduction
+edge/face manually. The trade-off in our case is small for static-shape
+pairs (boxes don't have many support directions), but meaningful for
+mesh / smooth shapes where the reduction path matters.
+
+**meep EPA** (`gjk/expanding_polytope_algorithm.js`): up to 64 faces, 32
+loose edges, 64 iterations; degeneration is silently swallowed and the
+best-so-far face is returned. The `debugger;` statement at line 136 is a
+leftover that should be removed.
+
+**Parry EPA** (`parry/src/query/epa/epa3.rs`): adaptive `eps_tol` scaled
+by `max(scale_of_coordinates, 1.0)` to prevent spurious failures on
+large-coordinate geometry (fixes parry issue #415). Hard 100-iter cap.
+Both stagnation- and progress-based termination criteria.
+
+The **scale-adaptive tolerance** is a concrete, copyable improvement. Our
+`EPA_TOLERANCE = 0.0001` (line 10) is an absolute threshold — on a
+character-scale shape it's fine but on a kilometre-scale heightmap
+collision it's well below the practical noise floor and may cause early
+"convergence" on a degenerate face.
+
+### 2.4 MPR
+
+Both engines have MPR (`gjk/mpr.js` vs `parry/src/query/contact/`/various).
+Parry uses MPR primarily for **support-map vs support-map proximity tests**
+where GJK is overkill. Our `mpr.js` is used only by `shape_cast.js:364` for
+the normal-recovery probe — it's **not wired into `narrowphase_step.js`** as
+a fallback for non-convergent EPA on smooth shapes (PLAN.md notes this as
+its intended role). That's a low-risk swap: per-pair config to prefer MPR
+for shapes whose support function is smooth (sphere, capsule, possibly
+the unwired cylinder), or as a fallback when EPA returns a non-positive
+depth.
+
+### 2.5 Manifold caching and contact persistence
+
+This is where meep and Parry diverge most subtly.
+
+- **meep** (`contact/ManifoldStore.js`):
+  - Slot keyed by `(idA, idB)` via `PairUint32Map`.
+  - Per-slot meta: count, touched bit, prev_touched bit, grace counter.
+  - Each step `acquire(idA, idB)` flips touched=1. `advance_frame` rolls
+    touched → prev_touched.
+  - Contacts overwritten each frame (`clear_contacts(slot)` at
+    `narrowphase_step.js:687`); only `j_n`, `j_t1`, `j_t2` (warm-start
+    impulses) survive into the next frame.
+  - **There is no per-contact feature-id matching.** Warm-start impulses
+    are written by index `(slot, k)` — if the narrowphase emits contacts
+    in a different order this frame vs. last frame (which can absolutely
+    happen for a rotating box-on-box), the warm-start impulses are applied
+    to the **wrong contact points**.
+
+- **Parry** (`contact_manifold.rs`):
+  ```rust
+  pub struct TrackedContact<Data> {
+      pub local_p1: Vector,
+      pub local_p2: Vector,
+      pub dist: Real,
+      pub fid1: PackedFeatureId,  // feature id on shape 1
+      pub fid2: PackedFeatureId,  // feature id on shape 2
+      pub data: Data,             // user data (warm-start impulses)
+  }
+  ```
+  Each per-pair contact generator (cuboid-cuboid, cuboid-triangle, etc.)
+  produces contacts with **stable feature IDs**: e.g. `face(2) × face(5)`,
+  `edge(0,1) × vertex(3)`, etc. `match_contacts()` then transfers the
+  user data (impulses) onto the corresponding contact in the new
+  manifold by feature-ID equality.
+
+**This is a real correctness bug in our solver.** The warm-start cache
+arithmetic is fine in isolation, but in the rotating-box-on-box case the
+reduction pass in `box_box_manifold.js:498-530` and the candidate
+reduction in `narrowphase_step.js:149-178` produce contact orderings that
+can re-shuffle between frames. The solver then warm-starts contact[0]
+with impulse from a different physical contact, smearing across the
+manifold and producing the kind of low-amplitude jitter that's commonly
+attributed to "PGS being lossy" but is actually warm-start mis-aim.
+
+There are two ways to fix:
+
+1. **Add feature-id columns to the manifold stride**, populate from
+   `box_box_manifold` (the reference + incident face indices + Sutherland-
+   Hodgman edge index already define a feature ID), and match by
+   feature ID when assigning impulses. Parry's `match_contacts` is the
+   reference. Stride goes from 13 → 15 floats. Needs cooperation from
+   each closed-form pair generator.
+2. **Fallback: match by spatial proximity** like Parry's
+   `match_contacts_using_positions`. Cheaper to implement (just a 4×4
+   distance matrix at solver pre-step) but loses precision when contacts
+   are within the matching threshold of each other.
+
+This is the single most impactful **correctness** fix in this review.
+
+### 2.6 Solver: PGS vs PGS-S, Baumgarte vs split-impulse
+
+Rapier carries **two impulses and two RHS values** per constraint element:
+
+```rust
+pub struct ContactConstraintNormalPart<N> {
+    torque_dir1, torque_dir2, ii_torque_dir1, ii_torque_dir2,
+    rhs,             // restitution + position bias
+    rhs_wo_bias,     // pure non-penetration
+    impulse,         // this substep's accumulator
+    impulse_accumulator,  // cumulative across substeps
+    r, r_mat_elts,
+}
+```
+
+Inside a substep:
+
+```
+for _ in 0..num_internal_pgs_iterations:
+    contact_constraints.solve(...)        // uses rhs
+for _ in 0..num_internal_stabilization_iterations:
+    contact_constraints.solve_wo_bias(...) // uses rhs_wo_bias
+```
+
+The two-phase scheme is exactly what defuses the failure modes PLAN.md
+documents (restitution × warm-start, Baumgarte K× amplification):
+
+- Restitution is part of `rhs`, applied once during the bias phase. The
+  no-bias stabilisation pass operates on `rhs_wo_bias` so the
+  restitution impulse cannot be subtracted away by a later iteration
+  seeing "separating velocity".
+- Position correction is part of `rhs`, applied during the bias phase, and
+  the no-bias pass cleans up the post-bias velocity contamination.
+- Forces are integrated once per substep into `solver_vels_increment`
+  outside the iteration loop, so the accumulator-timing issue doesn't
+  arise (each substep gets the right share of the force).
+
+Our `solver/solve_contacts.js:531` does:
+
+```js
+const lambda_n = -m_eff_n * (vn + bias_n);
+const sum_n = j_n_accum + lambda_n;
+const new_j_n = sum_n > 0 ? sum_n : 0;  // CLAMP
+```
+
+with `bias_n` blending Baumgarte position correction and restitution
+(`solve_contacts.js:442-457`). On the very next iteration the same
+contact sees a different `vn`, recomputes the full `(vn + bias_n)` target,
+and can subtract whatever the previous iteration added. Restitution is
+particularly fragile because it's a one-shot impulse that should be
+applied during the contact's *first* solver pass and never adjusted
+again, but our scheme keeps re-targeting it.
+
+**This is the architectural lock that prevents TGS**, as `PLAN.md` documents.
+The minimum fix is to **split `rhs` into `rhs_pos` and `rhs_restitution` +
+`rhs_velocity_only`**, apply the position+restitution bias only on a
+single dedicated iteration (or set of iterations), and then run the rest
+of the solver on the velocity-only target. Even without TGS substepping,
+this would defuse the warm-start × restitution interaction documented in
+`solve_contacts.js:452-456`.
+
+### 2.7 CCD
+
+Rapier has full CCD (`ccd/ccd_solver.rs`):
+
+```rust
+pub struct CCDSolver;   // stateless
+```
+
+with three phases:
+1. `update_ccd_active_flags()` based on velocity threshold.
+2. `predict_impacts_at_next_positions()` — sweeps a body's collider
+   against the broadphase, finds first TOI.
+3. `clamp_motions()` — clamps motion to stop just before impact.
+
+We have speculative-margin-via-fat-AABB only. PLAN.md flags this as a
+backlog item; the falling-tower bench is the concrete failure case (180
+of 1000 bodies tunnel through a 1 cm floor on a 1 km drop).
+
+The shape-cast machinery already in `queries/shape_cast.js` provides the
+TOI primitive. The missing piece is the per-body opt-in CCD path that
+runs after broadphase but before solver: identify high-speed bodies
+(`|v| · dt > min_extent`), sweep their colliders, clamp dt accordingly.
+This is conceptually parallel to `__wake_pairs` — a stage between
+broadphase and narrowphase.
+
+### 2.8 Concave / triangle mesh
+
+Both engines decompose a non-convex shape into triangles overlapping the
+convex's AABB, then run a per-triangle algorithm:
+
+- **meep** (`narrowphase_step.js:368-577`): generic per-triangle GJK+EPA
+  with a one-sided face-normal rejection. The `Triangle3D` support
+  function is degenerate along the face-normal axis (all 3 verts project
+  to the same value), so EPA can't tighten near vertical contacts — the
+  exact failure documented in PLAN.md's known limitations.
+
+- **Parry** (`contact_manifolds_trimesh_shape.rs`): dispatches the
+  per-triangle contact to the appropriate **closed-form** generator
+  (`contact_manifolds_cuboid_triangle.rs`, `contact_manifolds_convex_ball.rs`,
+  etc.). Triangle is treated as a polygonal feature with explicit face,
+  3 edges, 3 vertices and **stable feature IDs** per sub-feature.
+
+The actionable port is Parry's `contact_manifolds_cuboid_triangle.rs` —
+SAT + face clipping — adapted to our `Triangle3D` flyweight and
+emitting into our `(world_a, world_b, normal, depth)` stride. A
+ball-vs-triangle closed form (Voronoi-region-based) is even simpler:
+clamp the sphere centre to the triangle and dispatch on which feature
+(face / edge / vertex) the clamp landed on.
+
+### 2.9 Queries (raycast, shape-cast, overlap)
+
+Parry's `QueryPipeline` mirrors what we have in `queries/`:
+
+| meep | Parry equivalent | Notes |
+|---|---|---|
+| `raycast.js` | `query/ray/...` | meep is broadphase AABB-hit only; PLAN.md flags the narrowphase refinement as TODO |
+| `shape_cast.js` | `shape_cast_support_map_support_map.rs` | meep uses **bisection on GJK overlap**; Parry uses **GJK directional_distance / minkowski_ray_cast** — conservative advancement |
+| `overlap_shape.js` | `query/intersection_test/...` | both use broadphase + GJK |
+
+**The shape_cast approach is genuinely different.** Our bisection is
+robust and works, but is O(`log₂(tMax / tol)`) GJK calls per candidate.
+Parry's GJK directional_distance is **continuous refinement** — it
+advances the ray origin in the Minkowski difference each iteration and
+tightens both bounds simultaneously. Typically converges in 5-15
+iterations vs. our 32-step coarse linear search + up to 32 bisection
+steps. This is a clean port target since we already have all the
+support-function machinery.
+
+### 2.10 Determinism
+
+- **Rapier** advertises cross-platform determinism with f32 by avoiding
+  transcendentals and relying on IEEE-754 round-to-nearest behaviour on
+  the supported platforms. Their integration uses `mul_add`-free math
+  where possible.
+- **meep** is **same-runtime bit-exact** because `Math.sin/cos/exp/log` are
+  ULP-correct but not bit-exact across V8 / SpiderMonkey / JSC. PLAN.md
+  is explicit about this. The codebase's typed-array discipline (direct
+  index writes, no `Vector3.set` on hot paths, min-heap free list, sorted
+  awake-list iteration, canonical pair ordering) is the right
+  infrastructure.
+
+The action item to close the cross-runtime gap is documented in PLAN.md
+(soft-float library for transcendentals). Not a gap in scope of this
+review.
+
+---
+
+## 3. In-depth correctness comparison on six key touchpoints
+
+### 3.1 Warm-start impulse correctness across reordered contacts
+
+**meep** `solver/solve_contacts.js:467-476`:
+
+```js
+const j_n  = data[off + 10];
+const j_t1 = data[off + 11];
+const j_t2 = data[off + 12];
+const Px = nx * j_n + t1x * j_t1 + t2x * j_t2;
+// ... apply ...
+```
+
+The warm-start impulses live at fixed slot offsets `(slot, k)` in the
+manifold cache. The new frame's narrowphase calls `manifolds.clear_contacts(slot)`
+(at `narrowphase_step.js:687`), which **zeroes the data array including
+`j_n`, `j_t1`, `j_t2`** (see `ManifoldStore.js:233-239`):
+
+```js
+clear_contacts(slot) {
+    const meta_off = slot * SLOT_META_STRIDE;
+    this.__meta[meta_off + 2] = this.__meta[meta_off + 2] & ~COUNT_MASK;
+    const data_off = slot * SLOT_DATA_STRIDE;
+    this.__data.fill(0, data_off, data_off + SLOT_DATA_STRIDE);
+}
+```
+
+**Bug:** the docstring at `ManifoldStore.js:230` claims "narrowphase
+determines the pair is separated this frame" but `narrowphase_step.js:687`
+calls it **unconditionally** at the start of every narrowphase pass for
+every pair, before the new contacts are computed. So warm-start
+impulses are actively **destroyed on every frame**, every contact, every
+pair. That's exactly the opposite of what warm-starting requires.
+
+Walk-through:
+- Frame N: pair has contact[0] with `j_n = 5.2`. End of frame: solver has
+  applied impulses, data[10] = 5.2.
+- Frame N+1: `acquire()` is a no-op (slot exists, just sets touched).
+- Frame N+1 narrowphase: `clear_contacts(slot)` runs — **5.2 is wiped to 0**.
+- Frame N+1 contacts written via `set_contact`, which **does not touch
+  data[10..12]** (per its own docstring at `ManifoldStore.js:270` "warm-start
+  impulses are preserved") — but they were already zeroed by `clear_contacts`.
+- Solver pre-step at `solve_contacts.js:468-470` reads `j_n=0, j_t1=0, j_t2=0` →
+  no warm-start impulse applied.
+
+**The warm-start cache is effectively cold every frame.** This is a real
+bug — the comment at `ManifoldStore.js:269` "warm-start impulses are
+preserved across calls" describes intended behaviour that the
+`clear_contacts → set_contact` sequence in narrowphase actively breaks.
+
+Even if `clear_contacts` were fixed to preserve `j_n/j_t1/j_t2`, the
+**ordering instability** from §2.5 is the deeper issue: contacts[0]
+this frame is not contact[0] last frame for any non-trivial pair. Need
+feature-ID or position matching.
+
+**Improvement opportunity:**
+1. Make `clear_contacts` either skip the j_n/j_t1/j_t2 stride elements, or
+   accept a flag for "narrowphase contact replacement" vs. "pair separated".
+2. Implement feature-id or position matching at the contact-write site
+   (`narrowphase_step.js:691-700`) — when writing the new contact, find
+   the most-similar previous contact in this slot and **inherit its
+   warm-start impulses**.
+
+### 3.2 Box-box manifold reduction non-determinism vs. Parry's clipping
+
+**meep** `box_box_manifold.js:486-531`: reduce candidates to 4 by
+"deepest, then max-distance-to-kept-set" — a perimeter-maximisation
+heuristic. The ordering depends on numerical tiebreaks in
+`d2 < min_d2` comparisons. For symmetric inputs (two axis-aligned
+unit cubes at integer positions), this can produce different orderings
+in two otherwise-identical PhysicsSystem instances if floating-point
+arithmetic happens to differ by 1 ULP somewhere upstream.
+
+**Parry** `contact_manifolds_cuboid_cuboid.rs`: clipping produces
+**at most 4** candidates by construction (face × face Sutherland-Hodgman
+on a rectangle gives ≤ 8 corners; the four "deepest" are kept by a
+**feature-ID-based selection**, not a metric). Stable orderings.
+
+Our reduction also doesn't guarantee that the **deepest 4** are kept —
+it keeps deepest #1, then maximises spread. A cluster of three deep
+contacts plus one shallow but spread-out contact will lose one of the
+deep ones. For a flat box-on-floor with one corner slightly higher,
+that's the right behaviour. For a tilted box where the corner contacts
+matter, it can drop important contacts.
+
+**Improvement opportunity:** mirror Parry's feature-ID-based 4-of-8 pick
+(each corner of the clipped polygon carries the source edge and incident
+corner id; pick the four corners that span the largest convex hull).
+
+### 3.3 EPA on smooth supports — confirmed degeneration
+
+**meep** `expanding_polytope_algorithm.js:255-263`:
+
+```js
+if (num_loose_edges >= EPA_MAX_NUM_LOOSE_EDGES) {
+    // Polytope degenerated ...
+    // Bail out and return the closest-face result accumulated so far ...
+    break;
+}
+```
+
+The fallback path returns the **last closest face's normal × `dot_p_search_dir`**
+as the depth — but `dot_p_search_dir` was computed from a new support point
+that the polytope-update loop just **discarded** (because it can't fit it
+into the polytope without exceeding the cap). That's a mismatch: the depth
+is the projection of a vertex that isn't on the polytope, while the
+direction is the polytope's closest-face normal. The two should be consistent.
+
+**Parry** `epa3.rs`: scale-adaptive tolerance + 100-iteration cap; when
+it bails, it returns the closest face's barycentric-projected point, which
+is by construction a Minkowski-difference point. Output is a true MTV.
+
+**Improvement opportunity:**
+1. On the EPA bail path, return `closest_face_normal * min_dist` (the
+   stored `min_dist` for the chosen closest face), not
+   `normal * dot_p_search_dir`. `min_dist` is the actual signed distance
+   from origin to the closest face plane, which is the right MTV magnitude.
+2. Scale `EPA_TOLERANCE` by `max(scale_of_coordinates, 1.0)` — same fix
+   Parry applied for issue #415. The coordinate scale is the max of the
+   incoming simplex's coordinate magnitudes.
+3. Remove the leftover `debugger;` at `expanding_polytope_algorithm.js:136`.
+
+### 3.4 Concave dispatch: face-normal rejection + concave-vs-concave
+
+**meep** `narrowphase_step.js:519`:
+
+```js
+if (ex * fnx_w + ey * fny_w + ez * fnz_w <= 0) continue;
+```
+
+This is the "convex on outward side of triangle" test. **Correctness
+concern:** for a sphere centred slightly above a flat heightmap, the
+sphere-vs-triangle GJK+EPA may report an MTV that's near-perpendicular to
+the face (good — `fnx_w · ex > 0`), but for cases where the sphere is
+overlapping two adjacent triangles (one slanted, one horizontal), the
+sphere-vs-slanted-triangle's EPA can return an MTV that points *into*
+the slanted face. The rejection then skips that triangle, leaving only
+the horizontal-face contact — which is correct here. But:
+
+The face-normal rejection runs **after** sign-canonicalisation against the
+triangle centroid → convex centre vector. For a thin shape (capsule
+oriented along the slope), the centroid-to-centre vector might be near-
+parallel to the face, and the sign canonicalisation flips inconsistently
+between adjacent triangles, producing alternating contact directions
+across the heightmap. The `compute_penetration.js` half-space pre-test
+sidesteps this; the narrowphase doesn't.
+
+**Improvement opportunity:** for primitive-vs-triangle, use the
+half-space approach (project sphere centre / box corner / capsule
+endpoints onto the triangle plane) like `compute_penetration.js` does.
+This is roughly half the work of a closed-form ball-vs-triangle and
+already validated in our codebase. The full closed-form (§2.8) is
+strictly better — see backlog.
+
+### 3.5 GJK iteration tolerance and convergence detection
+
+**meep** `gjk.js:96-128`: 64-iter cap, **no progress check between
+iterations**. The only termination is "origin enclosed in tetrahedron"
+(return true) or "new support doesn't pass the origin" (return false).
+Non-progressing inputs (degenerate Minkowski difference at exactly
+touching, smooth-on-smooth tangent contact) run to 64 iterations and
+return false.
+
+**Parry** `gjk.rs`: progress check — if the new support's signed distance
+to the previous closest-feature plane improves by less than `eps`, declare
+convergence. Faster termination on near-touching pairs (sphere tangent to
+sphere); same false return on actual non-overlap but with fewer
+iterations.
+
+**Improvement opportunity:** add a progress check on the new support
+point's projection along the search direction. If `dot(new_support, dir)`
+is within tolerance of the previous iteration's value, declare convergence.
+On the practical level this affects performance (smooth-on-smooth) more
+than correctness (we already return false on the iteration cap), but
+falling out earlier on tangent contact is a real perf+stability win for
+the broadphase-edge case.
+
+### 3.6 Shape-cast bisection vs. GJK conservative advancement
+
+**meep** `queries/shape_cast.js:267-297`: AABB slab interval narrowing +
+32-step coarse linear search + up to 32 bisection steps + MPR normal
+recovery.
+
+**Parry** `shape_cast_support_map_support_map.rs` → `gjk::directional_distance`:
+**continuous refinement** in Minkowski space, simultaneously tightening
+both bounds. Converges in ~5-15 iterations on typical inputs because each
+iteration uses geometric structure (the support hyperplane at the current
+ray position) rather than blind bisection.
+
+Trade-offs:
+- Bisection is **simpler and dead-reliable** — no degenerate cases.
+- Conservative advancement is **substantially fewer GJK calls** —
+  Parry's typical TOI query is in the 10s of microseconds; ours is
+  64-step bisection per candidate.
+
+For our workload (kinematic character controllers — a few queries per
+frame), the perf delta is small. **Not a high-priority change** but
+worth knowing about. The normal recovery via MPR probe-step is fine
+(`shape_cast.js:350-381`); Parry's TOI returns the contact normal as
+part of the same conservative-advancement loop, so there's no separate
+probe needed.
+
+### 3.7 Friction: tangent disk vs. Coulomb pyramid
+
+**meep** `solver/friction_cone.js` (called from `solve_contacts.js:569`):
+two-tangent disk clamp — limits `sqrt(t1² + t2²) ≤ μ·j_n`. Coulomb cone
+in the velocity-impulse plane.
+
+**Parry**: looks similar in the `ContactConstraintTangentPart::solve` —
+Vector2 impulse clamped against `limit = mu * normal_part.impulse`.
+Functionally the same.
+
+No actionable difference; mark as parity.
+
+### 3.8 Body-centre application point fallback in GJK+EPA path
+
+**meep** `narrowphase_step.js:640-644`:
+
+```js
+return append_contact(count,
+    trA.position.x, trA.position.y, trA.position.z,
+    trB.position.x, trB.position.y, trB.position.z,
+    nx, ny, nz,
+    depth);
+```
+
+The fallback contact application points are **body centres**, with the
+detailed explanation in the inline comment (`narrowphase_step.js:623-639`)
+about flat-face support being multi-valued. This is correct in spirit but
+not in detail: for a tilted box on a flat floor with the GJK+EPA path
+(e.g. a non-cube convex hull on a heightmap triangle), the contact
+"point" sitting at the floor's centre rather than under the box produces
+the wrong torque arm. Specifically `r × n` is large and in the wrong
+direction. The Baumgarte position correction gets dampened by the large
+effective mass denominator and the box separates slowly.
+
+**Parry** computes contact points as **support-function witnesses**: for
+each shape, the support point in the contact normal direction is the
+contact application point. For flat-face shapes the witness is one of
+the (possibly-multi-valued) corners, and Parry deals with the
+multi-valued case by **picking a face when the support is multi-valued
+in the support direction**. Their `support_face` returns a
+`PolygonalFeature` rather than a point.
+
+**Improvement opportunity:** in the GJK+EPA fallback path, compute the
+contact point as `(posed_a.support(epa_dir) + posed_b.support(-epa_dir)) / 2`
+**modulo a check for flat-face support**: if `|support_a - centre_a|` or
+`|support_b - centre_b|` is suspiciously large relative to the shape's
+characteristic length, fall back to body centres. This narrows the body-
+centre fallback to its actually-correct domain (flat-faced shapes where
+the support is multi-valued) instead of using it for everything.
+
+---
+
+## 4. Simplicity and uniformity
+
+### 4.1 What's clean
+
+- **`BodyStorage`** (`body/BodyStorage.js`): SoA, generation packing,
+  deterministic free-list. Every primitive operation is one block of
+  index arithmetic. Hard to make simpler without losing functionality.
+- **`PairUint32Map`**: Robin Hood hashing in a single typed-array. No
+  allocation in steady state.
+- **`union_find.js`**: path halving in 8 lines. Minimal code.
+- **`integrate_velocity.js`**: 80 lines, no surprises, well-commented on
+  the implicit-decay damping recipe.
+- **`compute_fat_world_aabb.js`**: 60 lines, single-purpose, all constants
+  named.
+
+### 4.2 What's specialised in ways that earn their keep
+
+- **The two-BVH split** (`PhysicsSystem.staticBvh` + `dynamicBvh`): cheap
+  cognitively, real perf win for the "mostly-static world" target. Static
+  tree is built once and never touched.
+- **Atomic-island sleep with circular doubly-linked sleep groups**: the
+  alternative ("rebuild from scratch") would be much simpler but loses
+  the property that a hit at the base of a 100-stack wakes the whole
+  stack in one frame.
+- **Per-island CSR layout from `IslandBuilder`**: enables solver,
+  sleep-test, and event-dispatch to all iterate the same data structure.
+  Genuine reuse.
+- **`Triangle3D` flyweight + `triangle_buffer` reuse in the concave
+  dispatch**: zero allocation per triangle. The alternative (allocating
+  triangle instances) would be a hotspot.
+
+### 4.3 What's specialised in ways that don't earn their keep
+
+- **The narrowphase dispatcher** in `narrowphase_step.js:192-353` is a big
+  if-else chain on `isUnitSphereShape3D`, `isBoxShape3D`, etc., with
+  per-pair flipping of A/B identity (`isSphereA ? trA : trB` etc.). It's
+  legible but doesn't scale gracefully — adding cylinder support is
+  another ~50 lines of switch arms. Parry's approach is a **shape-type
+  pair → dispatch function** lookup, which is one indirection but
+  trivially extensible. A small registry table indexed by
+  `(shape_a_type << 8) | shape_b_type` would replace the entire chain
+  with one indirection. (NB: the type-tag identification via
+  `isUnitSphereShape3D === true` works in our JS world but the registry
+  generalises trivially.)
+- **Quaternion conjugate-rotate-conjugate boilerplate** in 5+ files
+  (`sphere_box_contact.js:48-57`, `capsule_contacts.js:39-71`,
+  `narrowphase_step.js:467-475`, `PosedShape3D.js:67-93`). Each
+  reimplements the same `v_world = q · v_local · q*` arithmetic inline.
+  Extracting a single `quaternion_apply` and trusting V8 to inline it
+  would clean the code at no cost. Parry has `Isometry` for exactly this.
+- **EPA's `loose_edges` array** (`expanding_polytope_algorithm.js:41`):
+  the `copyWithin` based edge removal scales poorly past ~32 edges and
+  the wrapping logic is fragile. Parry's adjacency-graph EPA is harder
+  to write but more robust. Worth keeping the current implementation
+  until it bites.
+
+### 4.4 Adding a new shape pair / constraint / query
+
+- **New shape pair (e.g. cylinder-box):** add a closed-form handler in
+  `narrowphase/cylinder_box_contact.js`, add an `isCylinderShape3D` type
+  tag, add a branch in `narrowphase_step.js` dispatch. ~3 files,
+  ~100-200 lines, mechanical. Parry's equivalent is also mechanical but
+  more strongly typed.
+- **New constraint (e.g. distance joint):** would need to thread `joints`
+  through the island builder + solver. PLAN.md notes "the solver loop is
+  already set up to iterate contacts ∪ joints; only constraint pre-step +
+  warm-start hook is missing." This is structurally easier than Parry
+  because we already iterate per-island.
+- **New query (e.g. sphere-cast):** `queries/sphere_cast.js` next to
+  `shape_cast.js` — but `shape_cast` is already generic, so this would
+  be a thin wrapper. Parry has both `cast_ray` and `cast_shape` for the
+  same reason.
+
+### 4.5 Recommended structural simplifications
+
+1. **Extract a `quaternion_apply(out, off, vx, vy, vz, qx, qy, qz, qw)`
+   helper** in `core/geom/3d/quaternion/`. Replace every inline copy in
+   the narrowphase.
+2. **Replace the narrowphase dispatcher's if-else chain with a small
+   typed dispatch table.** Adding shapes becomes a one-line entry.
+3. **Move warm-start impulse persistence out of the
+   `clear_contacts` path** (§3.1) — narrowphase calls a different method
+   `replace_contacts(slot, new_count)` that leaves j_n/j_t1/j_t2 alone,
+   and `clear_contacts` continues to be the "this pair has separated"
+   reset.
+4. **Centralise contact application-point selection** (§3.8) — a single
+   `contact_point_for_pair(shapeA, shapeB, epa_dir)` that knows when to
+   use witnesses vs. body centres, rather than scattering the heuristic
+   across narrowphase paths.
+
+---
+
+## 5. Headline findings, ranked
+
+1. **Warm-start cache is effectively cold every frame** (§3.1) due to
+   `clear_contacts(slot)` running unconditionally at the start of every
+   narrowphase pass. Real bug. Cheap fix.
+2. **Contact persistence lacks feature-id matching** (§2.5). When the
+   narrowphase emits manifold contacts in a different order this frame
+   vs. last, warm-start impulses apply to wrong contact points. Parry's
+   `TrackedContact { fid1, fid2, data }` is the model.
+3. **Solver is structurally locked out of TGS by combined Baumgarte+
+   restitution bias in `rhs`** (§2.6). The Rapier `rhs` vs `rhs_wo_bias`
+   split + impulse_accumulator architecture defuses the three failure
+   modes documented in PLAN.md. Significant work but unblocks the
+   biggest performance roadmap item.
+4. **Closed-form primitive-vs-triangle missing** (§2.8). PLAN.md flags as
+   the single biggest accuracy gap. Parry's `contact_manifolds_cuboid_triangle.rs`
+   + a Voronoi-region ball-vs-triangle are the port targets.
+5. **EPA scale-adaptive tolerance and consistent fallback** (§3.3). The
+   `EPA_TOLERANCE = 0.0001` is an absolute constant; should scale by
+   coordinate magnitude (Parry issue #415's fix). On the EPA bail path,
+   return `min_dist * face_normal`, not the inconsistent
+   `dot_p_search_dir * face_normal`. Also drop the leftover `debugger;`.
+6. **Concave-side per-triangle GJK+EPA cannot converge on degenerate
+   `Triangle3D` support** (§3.4). Already documented in PLAN.md. The
+   sphere-vs-triangle closed-form is the smallest first step (Voronoi
+   region clamp) and removes a class of skipped tests.
+7. **MPR is not wired into the narrowphase as an EPA fallback**.
+   PLAN.md correctly identifies this — the swap is in `narrowphase_step.js:580-644`
+   (replace the EPA call with `if (mpr(...)) ... else if (epa(...))` or
+   prefer MPR on smooth-shape pairs). The code already exists.
+8. **Box-box reduction can drop deep contacts in favour of spread**
+   (§3.2). Mirror Parry's feature-ID-based 4-of-8 corner pick on the
+   clipped polygon.
+9. **GJK lacks progress-based convergence detection** (§3.5). Adds a
+   single dot-product check per iteration; speeds up tangent contacts
+   and provides earlier-out on degenerate Minkowski differences.
+10. **CCD is speculative-margin-only**. Already in backlog. Rapier's
+    `predict_impacts_at_next_positions → clamp_motions` two-phase TOI
+    sweep is the architecture; our `shape_cast.js` already has the TOI
+    primitive.
+
+---
+
+## Appendix A: file-by-file Rapier/Parry references
+
+| meep file | Rapier/Parry counterpart |
+|---|---|
+| `ecs/PhysicsSystem.js` | `rapier/src/pipeline/physics_pipeline.rs` |
+| `body/BodyStorage.js` | `rapier/src/dynamics/rigid_body_set.rs` |
+| `broadphase/generate_pairs.js` | `rapier/src/geometry/broad_phase_bvh.rs` |
+| `broadphase/compute_fat_world_aabb.js` | (inline in `broad_phase_bvh.rs` refit logic) |
+| `contact/ManifoldStore.js` | `parry/src/query/contact_manifolds/contact_manifold.rs` |
+| `events/diff_manifolds.js` | `rapier/src/pipeline/event_handler.rs` |
+| `island/IslandBuilder.js` | `rapier/src/dynamics/island_manager/manager.rs` |
+| `core/collection/union-find/union_find.js` | (Rapier uses DFS, not union-find) |
+| `gjk/gjk.js` | `parry/src/query/gjk/gjk.rs` (`VoronoiSimplex` based) |
+| `gjk/expanding_polytope_algorithm.js` | `parry/src/query/epa/epa3.rs` |
+| `gjk/mpr.js` | `parry/src/query/mpr/...` |
+| `narrowphase/sphere_sphere_contact.js` | `parry/src/query/contact/contact_ball_ball.rs` |
+| `narrowphase/sphere_box_contact.js` | `parry/src/query/contact_manifolds/contact_manifolds_convex_ball.rs` |
+| `narrowphase/capsule_contacts.js` | `parry/src/query/contact_manifolds/contact_manifolds_capsule_capsule.rs`, `..._cuboid_capsule.rs` |
+| `narrowphase/box_box_manifold.js` | `parry/src/query/contact_manifolds/contact_manifolds_cuboid_cuboid.rs` |
+| `narrowphase/narrowphase_step.js` (concave path) | `parry/src/query/contact_manifolds/contact_manifolds_trimesh_shape.rs` |
+| `narrowphase/decomposition/heightmap_enumerate_triangles.js` | `parry/src/query/contact_manifolds/contact_manifolds_heightfield_shape.rs` |
+| `narrowphase/compute_penetration.js` | `parry/src/query/contact/contact_shape_shape.rs` |
+| `solver/solve_contacts.js` | `rapier/src/dynamics/solver/velocity_solver.rs` + `rapier/src/dynamics/solver/contact_constraint/contact_constraint_element.rs` |
+| `solver/friction_cone.js` | (inline in `ContactConstraintTangentPart::solve`) |
+| `integration/integrate_velocity.js` | `rapier/src/dynamics/solver/velocity_solver.rs` (between-substep integration) |
+| `integration/integrate_position.js` | `rapier/src/dynamics/solver/velocity_solver.rs` `integrate_positions` |
+| `inertia/world_inverse_inertia.js` | `rapier/src/dynamics/solver/...` (inline in constraint setup) |
+| `queries/raycast.js` | `parry/src/query/ray/...` |
+| `queries/shape_cast.js` | `parry/src/query/shape_cast/shape_cast_support_map_support_map.rs` (uses GJK conservative advancement) |
+| `queries/overlap_shape.js` | `parry/src/query/intersection_test/...` |