@woosh/meep-engine 3.1.0 → 3.2.0

package/src/engine/physics/REVIEW_001_ACTION_PLAN.md

@@ -1,642 +1,642 @@
-# REVIEW_001 — Action Plan
-
-Plan derived from the consolidated executive summary of three independent
-deep-reviewer reports (Bullet / Jolt / Rapier — all on opus, each
-instructed to read both sides line-by-line and write to disk).
-
-Source reviews live alongside this file:
-
-- `BULLET_REVIEW.md`
-- `JOLT_REVIEW.md`
-- `RAPIER_REVIEW.md`
-
-Each priority below cites the convergent findings from those reports and
-breaks the work into concrete actionable units with verification
-criteria. Effort estimates match the executive summary's prioritisation
-table.
-
----
-
-## P0 — Correctness blockers (hours)
-
-### P0.1 — Fix warm-start wipe
-
-**Status**: LANDED, activated alongside P1.2.
-
-**Bug.** Every frame for every active manifold slot, `dispatch_pair`'s
-accumulator path calls `manifolds.clear_contacts(slot)`. `clear_contacts`
-at `ManifoldStore.js:233` does `this.__data.fill(0, data_off, data_off + SLOT_DATA_STRIDE)`
-— zeroing the FULL slot stride, including offsets 10/11/12 (`j_n`,
-`j_t1`, `j_t2`). `set_contact` at line 258 then only writes offsets 0..9
-(position / normal / depth), leaving the warm-start impulses at zero.
-The comment at line 270 — `// j_n, j_t1, j_t2 are warm-start; preserved
-across calls` — is true at the per-call level but contradicted by the
-upstream wipe. Solver's warm-start sees zero accumulated impulses every
-frame.
-
-**Fix path (infrastructure landed).** `ManifoldStore` now exposes two
-operations:
-
-1. `begin_refill(slot)`: resets the contact count to zero (so the
-   next `set_contact(slot, 0, …)` lands at index 0) but does NOT zero
-   the data slab. Warm-start impulses survive. **Implemented.**
-2. `clear_contacts(slot)` is unchanged (genuine eviction zeroes
-   everything — still used at line 672 when narrowphase determines
-   the pair has no colliders).
-
-**Activation blocked.** The narrowphase callsite at line 687
-intentionally still calls `clear_contacts`. We tried flipping it to
-`begin_refill` and ran the full physics spec suite — the 4-cube
-atomic-sleep test (`PhysicsSystem.spec.js`) fails because the stack
-**collapses** within ~3 simulated seconds. Diagnostic capture at tick
-794:
-
-| cube | y (expected) | y (actual) | vy           | sleep_timer |
-|------|--------------|------------|--------------|-------------|
-| 0    | 0.5          | 0.4996     | 0            | 0.504 → SLEEPING (singleton) |
-| 1    | 1.5          | 0.4983     | ~1e-9        | 0.083       |
-| 2    | 2.5          | 0.689      | +0.21        | 0           |
-| 3    | 3.5          | 0.707      | bouncing     | 0           |
-
-Cube 1 has sunk **1 m below** its proper resting position (interpenetrating
-cube 0 by a whole edge). The reviewer's assumption "slot ordering is
-usually preserved" is false: `reduce_candidates` (`narrowphase_step.js:104`)
-reorders contacts by depth + max-spread, so the cached `j_n` at slot
-index `i` next frame is applied to a **different geometric contact**
-than the one whose impulse converged last frame. Stale impulses applied
-to wrong contacts destabilise the solver immediately on stacks.
-
-**Conclusion: P0.1 was bundled with P1.2.** Both landed together: the
-`begin_refill` primitive replaced `clear_contacts` at the narrowphase
-refill callsite, while P1.2's match-and-merge pass ensures the cached
-impulses follow the right contact slot index across the per-frame
-reshuffle that `reduce_candidates` introduces. The 4-cube atomic-sleep
-test (which collapsed when `begin_refill` was activated without
-matching) now passes with warm-start active.
-
-**Verification (landed).**
-- `ManifoldStore.spec.js`: asserts `j_n` / `j_t1` / `j_t2` survive a
-  `begin_refill` + `set_contact` cycle; `clear_contacts` still wipes
-  them; multi-contact preservation works.
-- `PhysicsSystem.spec.js`: 4-cube atomic-sleep test, 16-cube short
-  window stack, and all dependent integration tests pass with warm
-  start active.
-
----
-
-### P0.2 — Remove `debugger;` from EPA hot path
-
-**Status**: confirmed at `gjk/expanding_polytope_algorithm.js:136`.
-
-A `debugger;` statement is sitting in the EPA expansion loop. Production
-ship-blocker. Likely a leftover from a prior debug session.
-
-**Fix.** Delete line 136 (and 135 / 137 if they're related guard /
-comment). Confirm by grepping `^\s*debugger` across the package — there
-should be zero hits afterward.
-
-**Verification.** `grep -rn "debugger" src/engine/physics/` returns
-nothing.
-
-**Effort**: minutes.
-
----
-
-## P1 — Largest accuracy/correctness wins (days)
-
-### P1.1 — Closed-form sphere / box / capsule-vs-triangle
-
-**Status**: LANDED in three sub-commits — P1.1a (sphere), P1.1b
-(box), P1.1c (capsule).
-
-**Problem.** The narrowphase concave dispatch (added in commit 28ef0e894)
-decomposes heightmaps / meshes to triangles and runs per-triangle GJK +
-EPA. `Triangle3D` has a degenerate support along its face-normal axis
-(all 3 vertices project to the same value), so GJK precision is poor —
-hence the skipped settle tests in `narrowphase_concave.spec.js` and the
-documented sink-through-heightmap behaviour. `compute_penetration`
-already works around this for sphere-vs-heightmap by using the
-half-space test on the face plane, but the narrowphase still uses
-per-triangle GJK+EPA.
-
-**Files to add** (suggested layout, paralleling existing
-`sphere_box_contact.js` / `capsule_contacts.js`):
-
-- `engine/physics/narrowphase/sphere_triangle_contact.js` ✅ LANDED —
-  closest-point on triangle (uses
-  `computeTriangleClosestPointToPointBarycentric`), distance to sphere
-  centre, depth = R − dist when overlap. Output stride matches
-  `sphere_box_contact`. Includes upfront degenerate-triangle guard
-  (zero-area triangle returns false) and dist=0 fallback to face
-  normal. 11 unit tests covering face, edge, vertex regions,
-  separation, tangent, singular, degenerate, tilted, indexing.
-- `engine/physics/narrowphase/box_triangle_contact.js` ✅ LANDED —
-  SAT over 13 axes (3 box face normals + 1 triangle normal + 9 edge-edge
-  crosses) with the corrected asymmetric `min(push_pos, push_neg)`
-  per-axis MTV (triangle projection is NOT symmetric around its
-  centroid, unlike box's). Three contact-generation branches:
-  - **Box face winner** → reference = box face, incident = triangle.
-    Project triangle into the face's (u, v) basis, clip with 4 axis-aligned
-    Sutherland-Hodgman passes, recover world contacts on the triangle
-    plane.
-  - **Triangle face winner** → reference = triangle, incident = box
-    face most antiparallel to contact normal. Project the box quad
-    onto the triangle plane (basis = AB normalised + tn × AB), clip
-    with 3 general-half-plane passes against the triangle's 3 edges,
-    recover world contacts on the box face plane.
-  - **Edge-cross winner** → identify the relevant box edge (one of 4
-    parallel edges, selected by sign of `-n` projected onto the two
-    perpendicular box axes) and triangle edge, find closest pair via
-    `line3_closest_points_segment_segment`, emit single contact.
-
-  12 unit tests covering separation, degenerate triangle, box face
-  winner (single + multi-point), triangle face winner, deep
-  penetration, +/-Y face flip, self-consistency reconstruction
-  (`tri_pt = box_pt + depth * n`), rotated box, near-corner case,
-  translated box, output indexing.
-- `engine/physics/narrowphase/capsule_triangle_contact.js` ✅ LANDED —
-  multi-point manifold via primary segment-vs-triangle closest-pair
-  + cap-centre sphere queries. The `segment_triangle_closest` helper
-  takes the minimum of 6 candidates: (case 0) segment crosses
-  triangle plane inside the face → distance 0; (cases 1, 2) each
-  segment endpoint vs `computeTriangleClosestPointToPointBarycentric`;
-  (cases 3, 4, 5) `line3_closest_points_segment_segment` against each
-  triangle edge. Primary contact emitted from the closest pair; each
-  cap centre then runs `sphere_triangle_contact` and is spatially
-  deduplicated against the primary. Result: 1..3 contacts.
-  Pattern matches `capsule_box_multi_contacts`. 10 unit tests.
-
-**Wiring.** Inside `narrowphase_step.js`'s concave branch (the one that
-calls `decompose_to_triangles`), per-triangle dispatch by convex
-shape type:
-
-```js
-if (convex_shape.isUnitSphereShape3D)   → per-triangle sphere_triangle_contact    [LANDED]
-else if (convex_shape.isBoxShape3D)     → per-triangle box_triangle_contact       [LANDED]
-else if (convex_shape.isCapsuleShape3D) → per-triangle capsule_triangle_contact   [LANDED]
-else                                     → existing per-triangle GJK+EPA fallback
-```
-
-Each closed-form path reads the triangle's three vertices from the
-decomposition buffer at stride 10, rotates them to world space (3
-quat rotations per triangle — ~63 flops, negligible vs. the GJK + EPA
-cost being replaced), runs its solver, then runs the same one-sided
-face-normal rejection and normal-based dedup as the EPA path. The
-closed-form paths use ACTUAL surface witnesses (closest point on
-triangle, sphere surface point) rather than body centres — exact in
-both directions, unlike the EPA fallback where support witnesses can
-be arbitrarily far from the contact patch.
-
-**Verification (P1.1 fully landed).**
-- 11 unit tests in `sphere_triangle_contact.spec.js`.
-- 12 unit tests in `box_triangle_contact.spec.js`.
-- 10 unit tests in `capsule_triangle_contact.spec.js`.
-- The two sphere settle tests in `narrowphase_concave.spec.js`
-  un-skipped and passing.
-- The torus-knot test in `PhysicsSystem.spec.js` un-skipped and
-  passing.
-- Full physics suite: 686 passing (was 650 pre-P1.1a — delta:
-  +11 sphere_triangle + 12 box_triangle + 10 capsule_triangle unit
-  tests + 2 un-skipped sphere settle tests + 1 un-skipped torus-knot
-  test).
-
-**Optional follow-up** (not blocking).
-- Tighten `compute_penetration.spec.js`'s mesh-cube test (currently
-  documented as approximate due to closed-mesh over-reporting on side
-  faces). Note: `compute_penetration` is a separate code path from
-  `narrowphase_step.js` — it uses its own GJK+EPA / MPR pipeline and
-  the half-space workaround; landing the narrowphase fast-path
-  doesn't change it. Refactoring `compute_penetration` to share the
-  sphere / box / capsule-triangle fast-paths is a separate cleanup.
-
-**Effort**: LANDED.
-
----
-
-### P1.2 — Feature-id contact tracking across frames
-
-**Status**: LANDED.
-
-**Problem.** Even after the P0 warm-start wipe is fixed, the manifold
-slot reuses contact-index ordering across frames. Contact slice indices
-shift due to:
-
-- Depth-spread reduction in `reduce_candidates` (`narrowphase_step.js:104`)
-  — picks deepest, then max-spread; ordering depends on candidate
-  generation order which depends on shape-pair details.
-- Box-box face clipping reorders contacts based on which clipped
-  vertices survive (`box_box_manifold.js`).
-- The triangle decomposition emits cells in row-major order — a sphere
-  rolling along a heightmap sees different cell indices each frame.
-
-Reviewers cite two reference designs:
-
-1. **Parry's `TrackedContact { fid1, fid2 }`** — each contact carries
-   feature ids on both sides; the cache matches by `(fid1, fid2)` pair.
-2. **Jolt's `mContactPointPreserveLambdaMaxDistSq`** — match by
-   local-space position proximity (default 0.04 m squared).
-
-We're already laid out for (1): the decomposition pipeline emits stable
-feature ids per triangle, and the convex primitive narrowphase paths
-can derive feature ids from vertex / edge / face indices.
-
-**What landed.**
-
-1. `CONTACT_STRIDE` extended from 13 to 14 (added `feature_id` at
-   offset 13). `CANDIDATE_STRIDE` extended from 10 to 11 in
-   `narrowphase_step.js` (mirroring slot layout).
-2. `set_contact` gained an optional `feature_id` parameter (default 0
-   — preserves backwards-compatible call signatures).
-3. New `clear_impulses(slot, idx)` method on `ManifoldStore`: zeroes
-   only the j_n / j_t1 / j_t2 fields at one contact index. Used by
-   the match-and-merge pass when a new candidate has no matching
-   prev-frame contact.
-4. New `feature_id_of(slot, idx)` accessor for testability.
-5. Per-dispatch feature_id derivation in `narrowphase_step.js`:
-   - `sphere_sphere`: fid = 1 (single contact, real feature)
-   - `sphere_box`: fid = voronoi region 1..27 via new
-     `sphere_box_voronoi_fid` helper (inverse-rotates sphere centre
-     to box-local frame, buckets each component into {<-h, [-h,h], >h})
-   - `box_box`: fid = 0 (clipped-contact ordering isn't cross-frame
-     stable; position-fallback handles per-contact matching)
-   - `capsule_capsule`, `capsule_sphere`: fid = 1
-   - `capsule_box` multi-point: fid = k+1 (per-sub-contact index)
-   - Triangle (concave) path: fid = triangle's pre-existing
-     decomposition fid
-   - GJK+EPA fallback: fid = 0 (no good source)
-6. Match-and-merge pass replaces `clear_contacts(slot)` at the
-   narrowphase refill callsite:
-   - Snapshots prev contacts' `(feature_id, world_a, j_n, j_t1, j_t2)`
-     upfront into scratch (decouples read from write to avoid hazards
-     when the matching mapping shuffles indices)
-   - For each new candidate, finds best prev match — feature_id first
-     (only if both sides have fid ≠ 0), position-fallback within
-     `MATCH_TOL_SQR = 0.02 * 0.02` second
-   - Each prev contact can be claimed by at most one candidate
-     (`prev_claimed[]` flags)
-   - Calls `begin_refill(slot)` (count → 0, data preserved)
-   - Writes each new candidate at slot index k; copies matched prev
-     impulses to slot index k OR calls `clear_impulses(slot, k)` for
-     unmatched candidates
-7. Separation path (cand_count === 0) calls `clear_contacts(slot)` —
-   genuine eviction, drops stale impulses before contact is
-   re-established at a different feature.
-
-**Verification.**
-- `ManifoldStore.spec.js`: 5 new tests covering `set_contact`
-  feature_id round-trip, default fid = 0, impulse-preservation under
-  fid'd writes, `clear_impulses` scoping (one index, geometry +
-  feature_id intact), and per-index targeting.
-- `PhysicsSystem.spec.js`: 4-cube atomic-sleep test passes (was the
-  regression that gated P0.1 alone — confirms warm-start activation
-  is now stable for stacks).
-- Full physics test suite: 650 passing (was 645 pre-P1.2 — the
-  delta is the 5 new ManifoldStore tests; the 4-cube atomic-sleep
-  test which transiently regressed under P0.1-alone is now stable).
-
-**Effort**: landed.
-
----
-
-## P2 — Performance + precision (day-scale)
-
-### P2.1 — GJK separating-axis cache per manifold slot
-
-**Status**: LANDED (second attempt). The first attempt regressed basic
-settle tests and was reverted; this second pass landed the same
-optimisation via a different mechanism (no try/finally) in 6 inert /
-single-activation steps, each independently verified.
-
-**Mechanism — `gjk_core` + `gjk_with_axis`.** Instead of try/finally,
-gjk's body was extracted into `gjk_core(simplex, A, B, dir)` taking
-the search-direction buffer as a parameter. The existing
-`gjk(simplex, A, B)` is now a thin wrapper that sets the module's
-`scratch_dir` to `(1, 0, 0)` and delegates — semantically identical to
-the old code. The new `gjk_with_axis(simplex, A, B, axis_buf, axis_off)`
-creates a `subarray` view into `axis_buf[axis_off..+2]` and passes it
-to `gjk_core`; in-place writes to the view propagate to the caller's
-buffer automatically, so no try/finally writeback is needed. A guard
-catches zero / NaN / Inf seeds and resets to `(1, 0, 0)`.
-
-**Storage — `ManifoldStore.__slot_axis`.** A `Float64Array(capacity * 3)`
-parallel to the data buffer. Zeroed on slot `acquire` (so a recycled
-slot doesn't leak the previous pair's axis). Grows alongside the
-other slot arrays. Public accessors `slot_axis_buffer` and
-`slot_axis_offset(slot)`.
-
-**Activation — narrowphase.** The outer narrowphase loop computes
-`(gjk_axis_buf, gjk_axis_off)` from the slot and passes them through
-`dispatch_pair`. Both GJK + EPA fallback paths (body-level for
-non-closed-form convex pairs; per-triangle for non-(sphere/box/
-capsule) convex shapes against concave shapes) call `gjk_with_axis`
-with these arguments.
-
-**Sequencing — six steps with independent verification.** Each step
-either had no behaviour change (1, 3, 4) or was a small enough
-activation that a regression would be precisely revertable (5, 6).
-Verified at every step against the full physics suite:
-
-| Step | Change | Tests passing |
-|---|---|---|
-| 1 | Extract `gjk_core` from `gjk`; existing wrapper preserves semantics | 687 / 687 |
-| 2 | Add `gjk_with_axis` + 6 unit tests | 693 / 693 |
-| 3 | Add `ManifoldStore.__slot_axis` (dead code) | 693 / 693 |
-| 4 | Plumb axis args through `dispatch_pair` (dead code) | 693 / 693 |
-| 5 | Activate at the body-level GJK + EPA fallback | 693 / 693 |
-| 6 | Activate at the per-triangle EPA path in concave dispatch | 693 / 693 |
-
-**Why this attempt worked where the first didn't.** The first attempt
-wrapped the existing gjk body in try/finally to write back the cached
-axis on every return path. That should have been semantically
-identical for existing (cache-less) callers, yet basic settle tests
-regressed. Working theory: V8's JIT shape-inference handled the
-try/finally-wrapped function differently in ways that surfaced in the
-existing physics tests' numerical paths. The second attempt avoids
-try/finally entirely — `gjk_core` takes `dir` as a parameter, and
-writeback is automatic via shared memory (the caller's buffer IS the
-working buffer through a subarray view).
-
-**Improvement.** Bullet's `m_cachedSeparatingAxis` and Jolt's `ioV`
-in/out pattern store the last successful separation direction (or
-overlap-resolution direction) per manifold slot. Reuse next frame as
-GJK's initial direction. Reviewers cite 5–10× iteration reduction for
-quiescent persistent contacts — the common case in any settled stack.
-
-**Plan.**
-
-1. Add 3 floats to `SLOT_META_STRIDE` (or a parallel `Float32Array(slot_count * 3)`)
-   in `ManifoldStore` for the cached axis.
-2. Add an overload (or new entry point) to `gjk` accepting an initial
-   direction. Backward-compatible: existing callers (shape_cast,
-   compute_penetration, overlap) pass `null` and default to the
-   current `(1, 0, 0)`.
-3. In `narrowphase_step.js`'s `dispatch_pair`, before the GJK+EPA
-   fallback (line 312 area), read the cached axis from the slot; pass
-   it to GJK. After GJK terminates (with or without overlap), write
-   the final search direction back to the slot.
-
-**Verification.**
-- Microbench: count GJK iterations on the existing
-  `PhysicsSystem.bench.spec.js` torus-knot scene (currently skipped
-  for accuracy but useful here for iteration count). Expect 5–10×
-  reduction.
-- Determinism: same scene state → same iteration counts.
-
-**Effort**: 1 day.
-
----
-
-### P2.2 — Adaptive EPA tolerance + bail-out fix
-
-**Status**: PARTIALLY LANDED. Adaptive tolerance landed; bail-out
-magnitude consistency reverted (caused energy injection in closed-form
-dispatch paths that share the EPA depth metric — sphere-vs-sphere
-stack regressions launched bodies metres straight up).
-
-**Two issues.**
-
-1. `EPA_TOLERANCE = 0.0001` is absolute. For small-depth contacts (sub-mm)
-   this is comparable to the depth itself; EPA terminates with whatever
-   intermediate face it has, producing a noisy normal direction. Jolt
-   uses a relative tolerance scaled by current closest-face distance.
-2. The bail-out path at line 343 uses `dot_p_search_dir` where the
-   converged path uses `min_dist`. Inconsistent under iteration-cap
-   exit — the documented "best-effort fallback" returns a different
-   magnitude than the converged path of the same geometry.
-
-**Plan and outcome.**
-
-1. ✅ LANDED. Tolerance is relative: `EPA_TOLERANCE_REL = 1e-4`
-   floored by `EPA_TOLERANCE_ABS = 1e-6`. Terminate when
-   `dot_p_search_dir − min_dist < EPS_REL * max(|min_dist|, EPS_ABS)`.
-   This fixes EPA's failure-to-converge for sub-mm contacts where the
-   old absolute tolerance was larger than the depth itself.
-2. ❌ REVERTED. Swapping the convergence-path result magnitude from
-   `dot_p_search_dir` to `min_dist` (for consistency with the
-   iteration-cap bail-out path) was tried — and caused energy
-   injection in closed-form dispatch tests that share the EPA depth
-   metric (sphere-vs-sphere with restitution 0 launched a ball to 7
-   m). The convergence path retains its original magnitude. The
-   "inconsistency with bail-out" the reviewer noted is real but
-   benign in practice: bail-out fires on smooth-shape EPA failure
-   modes that the closed-form paths now bypass entirely (P1.1
-   sphere/box/capsule-vs-triangle); the consistency gap doesn't
-   surface anywhere observable.
-
-**Verification.**
-- EPA spec needs new cases at very small depth (sub-µm) confirming
-  termination is now stable and direction is well-defined.
-- The shape_cast bail-out fallback comparison should produce
-  consistent normals between converged and non-converged exits.
-
-**Effort**: ½ day. Less critical post-P0 because `shape_cast` and
-`compute_penetration` now use MPR for shallow-overlap normal recovery
-— this fix is for the GJK+EPA fallback in `narrowphase_step.js` only.
-
----
-
-## P3 — Mitigation improvements (half-day each)
-
-### P3.1 — Wire MPR as EPA fallback for mesh-involving pairs
-
-**Status**: LANDED.
-
-`gjk/mpr.js` was already implemented and tested; `shape_cast.js` and
-`compute_penetration.js` already used it. The two narrowphase EPA call
-sites — the per-triangle EPA path inside the concave dispatch (when
-the convex shape is not sphere/box/capsule) and the body-level
-GJK+EPA fallback — now both fall back to MPR when EPA returns
-zero / negative / NaN / Inf depth. MPR's portal-refinement convergence
-properties tend to succeed where EPA's face-expansion hits its
-iteration cap.
-
-**Plan.**
-
-1. In `dispatch_pair`'s GJK+EPA fallback, after EPA returns:
-   - If depth > 0 and direction passes the body-centre sanity check
-     → keep the EPA result.
-   - Else → re-run with MPR (same simplex inputs are not needed; MPR
-     starts from the PosedShape pair). If MPR returns true with a
-     positive depth, use that.
-2. For pairs involving a non-convex shape (heightmap/mesh dispatch via
-   the concave path), once the closed-form triangle solvers from P1.1
-   land, MPR fallback is only needed for the GJK+EPA leg of the
-   concave loop — same pattern.
-
-**Verification.**
-- The torus-knot test in `PhysicsSystem.spec.js` (currently skipped)
-  should make progress with the MPR fallback even before P1.1 lands.
-- `compute_penetration.spec.js`'s smooth-shape direction-precision
-  tests can tighten their tolerances.
-
-**Effort**: ½ day.
-
----
-
-### P3.2 — Box-box edge-edge closest-pair manifold
-
-**Status**: LANDED.
-
-In `box_box_manifold.js`'s edge-cross-axis-winner branch, the previous
-body-centre-midpoint fallback (both contact points collapsed to the
-midpoint of the two box centres — wrong lever arms, slow drift in
-skewed-box-stack scenarios) is replaced with a proper closest-pair
-computation:
-
-1. Decode `(i, j)` from `best_source = 6 + i*3 + j` — box A's local
-   axis index and box B's.
-2. Pick A's edge among 4 parallel edges along axis `i` by signing the
-   perpendicular axes against `-n` (the direction from A's centre
-   toward B). Same for B with `+n`.
-3. `line3_closest_points_segment_segment(out_st, edge_a, edge_b)`
-   returns `(s, t)`; reconstruct `contact_on_A = edge_a_p1 + s * (edge_a_p2 - edge_a_p1)`
-   and similarly for B.
-
-The "second contact for near-parallel edges" sub-improvement from the
-original plan turned out to be unnecessary in practice: when two box
-edges are near-parallel, the SAT axis `cross(edge_dir_a, edge_dir_b)`
-is degenerate (magnitude below `PARALLEL_EPS_SQR`) and the test_axis
-helper rejects it before it can win. Near-parallel-edge configurations
-end up with a face-axis SAT winner, which the face-clipping path
-already handles with multiple contacts.
-
-**Improvement.** When SAT identifies an edge-edge separating axis
-between two boxes (the cross-product axes in `box_box_manifold.js`),
-the current fallback emits a single contact at the midpoint of the
-edges' projections. The codebase already has
-`core/geom/3d/line/line3_closest_points_segment_segment.js` from a
-previous slice. Use it to emit the proper closest-pair, then evaluate
-spread for a second contact when the edges are near-parallel (within
-some `EDGE_PARALLEL_COS_TOL` like 0.999).
-
-**Plan.**
-
-1. In `box_box_manifold.js`'s edge-edge branch, replace the midpoint
-   contact with `line3_closest_points_segment_segment`'s output.
-2. Detect near-parallel edges via the cos-angle test on the edge
-   direction vectors; in that case emit a second contact at the other
-   endpoint of the parallel overlap interval. Standard SAT-with-clipping
-   pattern.
-
-**Verification.**
-- A box stacked at an angle on another box should reach steady state
-  without the slow drift that the single-midpoint contact causes.
-  Add a regression test alongside the existing box-box stack test.
-
-**Effort**: ½ day.
-
----
-
-## P4 — Strategic items (weeks; defer until pressed)
-
-### P4.1 — Split-impulse architecture → TGS substepping
-
-**Status**: PLAN.md documents this with the three interacting failure
-modes; reviewers confirmed the path forward in detail.
-
-**Recommendation.** Don't pursue until tall stacks become a frequent
-gameplay pain point. The current 4-cube atomic-sleep test passes; the
-16-cube short-window test catches manifold regressions. The pure-JS
-single-threaded budget is the more likely scaling wall — TGS doesn't
-help there.
-
-If/when we do it: Box2D-Lite-style split impulse — separate
-position-correction pseudo-velocity that doesn't contaminate real
-velocity, restitution as one-shot impulse at first contact (not a bias
-inside the loop), forces apply once at full dt before substeps.
-
-**Effort**: weeks.
-
-### P4.2 — Per-body linear CCD shape-cast
-
-**Status**: PLAN.md backlog ("Per-body linear CCD shape-cast").
-
-**Recommendation.** Already a focused future-work item with the 1km
-falling-tower reproducer (~180/1000 bodies tunnel). Worth scheduling
-once a gameplay use case demands it (e.g. fast projectiles, vehicle
-physics).
-
-**Effort**: week.
-
----
-
-## Not prioritised in the executive summary
-
-### Shape-cast bisection vs GJK conservative advancement
-
-Reviewers noted Parry's `gjk::directional_distance` (3–5× fewer GJK
-calls per query) but did not prioritise it. Our current bisection +
-slab-narrowing is functionally correct and acceptable; flag for revisit
-if `shape_cast` shows up in profiles. Effort would be a day.
-
----
-
-## Out of scope
-
-Confirmed from the consolidated review — these continue to be valid
-non-goals:
-
-- SIMD / WASM — V8 doesn't expose deterministic Float64x2 ops.
-- `SharedArrayBuffer` / multi-threaded solver — COOP/COEP headers
-  aren't always available.
-- Cross-runtime bit-exact determinism — would require a soft-float
-  library; same-runtime determinism covers single-device replay +
-  lockstep clients on matched JS engines.
-- Reduced-coordinate articulations — game-physics audience runs in
-  maximal coordinates by convention.
-
----
-
-## What's already strong (no action; for record)
-
-All three reviewers independently called these out as better than at
-least one reference engine. Useful confirmation that the design bets
-in PLAN.md paid off.
-
-- End-to-end SoA layout with generation-tracked stable IDs and
-  min-heap free list (deterministic ID reuse).
-- Atomic per-island sleep with chain wake via circular doubly-linked
-  `sleep_group_next` / `sleep_group_prev` — wake propagation
-  completes in one frame, vs Bullet/Jolt's bottom-up multi-frame
-  propagation.
-- Deterministic island roots (union-find with union-by-min-index) —
-  Rapier and Jolt do not strictly guarantee this.
-- Flat shape hierarchy, linear narrowphase dispatch,
-  monomorphic-when-hot `support()`.
-- Coulomb-cone disk friction clamp (true cone) — Bullet's default is
-  the anisotropic box clamp.
-- One-sided face-normal rejection on concave dispatch.
-
----
-
-## Recommended sequencing
-
-1. **P0.2 + P0.1 infra prep (LANDED)** — `debugger;` removed from EPA;
-   `ManifoldStore.begin_refill` added. Surgical, no behaviour change.
-   The initial attempt at full P0.1 activation (flipping the callsite
-   to `begin_refill`) was reverted after the 4-cube atomic-sleep test
-   caught a stack collapse, confirming the reviewer's "slot ordering
-   is usually preserved" assumption is false for stacks.
-2. **P1.2 + P0.1 activation (LANDED)** — feature-id contact tracking
-   and the match-and-merge pass. Warm-start now active and stable on
-   stacks; full physics suite green (650 passing).
-3. **P1.1 in three sub-commits — LANDED**:
-   - **P1.1a (sphere-vs-triangle).** `sphere_triangle_contact` module
-     + spec + narrowphase wiring; un-skipped two settle tests in
-     `narrowphase_concave.spec.js`.
-   - **P1.1b (box-vs-triangle).** `box_triangle_contact` module (SAT
-     + clipping for both face winners + closest-pair for edge-cross)
-     + spec + narrowphase wiring; un-skipped the torus-knot settle
-     test in `PhysicsSystem.spec.js`.
-   - **P1.1c (capsule-vs-triangle).** `capsule_triangle_contact`
-     module (segment-triangle closest-pair + cap-centre sphere
-     queries with spatial dedup) + spec + narrowphase wiring.
-4. **P2.1 (GJK axis cache) + P2.2 (EPA tolerance) together** — both
-   are localised, neither depends on the others.
-5. **P3.1 (MPR fallback) — LANDED.** Wired at both narrowphase EPA
-   call sites (per-triangle concave + body-level fallback). Behind
-   the existing depth-validity gate so it only fires when EPA fails.
-6. **P3.2 (box-box edge-edge) — LANDED.** `line3_closest_points_segment_segment`-based
-   contact-pair reconstruction; near-parallel edges already handled
-   by SAT degeneracy fallthrough to face axes.
-7. **P4 items**: defer until gameplay pressure.
+# REVIEW_001 — Action Plan
+
+Plan derived from the consolidated executive summary of three independent
+deep-reviewer reports (Bullet / Jolt / Rapier — all on opus, each
+instructed to read both sides line-by-line and write to disk).
+
+Source reviews live alongside this file:
+
+- `BULLET_REVIEW.md`
+- `JOLT_REVIEW.md`
+- `RAPIER_REVIEW.md`
+
+Each priority below cites the convergent findings from those reports and
+breaks the work into concrete actionable units with verification
+criteria. Effort estimates match the executive summary's prioritisation
+table.
+
+---
+
+## P0 — Correctness blockers (hours)
+
+### P0.1 — Fix warm-start wipe
+
+**Status**: LANDED, activated alongside P1.2.
+
+**Bug.** Every frame for every active manifold slot, `dispatch_pair`'s
+accumulator path calls `manifolds.clear_contacts(slot)`. `clear_contacts`
+at `ManifoldStore.js:233` does `this.__data.fill(0, data_off, data_off + SLOT_DATA_STRIDE)`
+— zeroing the FULL slot stride, including offsets 10/11/12 (`j_n`,
+`j_t1`, `j_t2`). `set_contact` at line 258 then only writes offsets 0..9
+(position / normal / depth), leaving the warm-start impulses at zero.
+The comment at line 270 — `// j_n, j_t1, j_t2 are warm-start; preserved
+across calls` — is true at the per-call level but contradicted by the
+upstream wipe. Solver's warm-start sees zero accumulated impulses every
+frame.
+
+**Fix path (infrastructure landed).** `ManifoldStore` now exposes two
+operations:
+
+1. `begin_refill(slot)`: resets the contact count to zero (so the
+   next `set_contact(slot, 0, …)` lands at index 0) but does NOT zero
+   the data slab. Warm-start impulses survive. **Implemented.**
+2. `clear_contacts(slot)` is unchanged (genuine eviction zeroes
+   everything — still used at line 672 when narrowphase determines
+   the pair has no colliders).
+
+**Activation blocked.** The narrowphase callsite at line 687
+intentionally still calls `clear_contacts`. We tried flipping it to
+`begin_refill` and ran the full physics spec suite — the 4-cube
+atomic-sleep test (`PhysicsSystem.spec.js`) fails because the stack
+**collapses** within ~3 simulated seconds. Diagnostic capture at tick
+794:
+
+| cube | y (expected) | y (actual) | vy           | sleep_timer |
+|------|--------------|------------|--------------|-------------|
+| 0    | 0.5          | 0.4996     | 0            | 0.504 → SLEEPING (singleton) |
+| 1    | 1.5          | 0.4983     | ~1e-9        | 0.083       |
+| 2    | 2.5          | 0.689      | +0.21        | 0           |
+| 3    | 3.5          | 0.707      | bouncing     | 0           |
+
+Cube 1 has sunk **1 m below** its proper resting position (interpenetrating
+cube 0 by a whole edge). The reviewer's assumption "slot ordering is
+usually preserved" is false: `reduce_candidates` (`narrowphase_step.js:104`)
+reorders contacts by depth + max-spread, so the cached `j_n` at slot
+index `i` next frame is applied to a **different geometric contact**
+than the one whose impulse converged last frame. Stale impulses applied
+to wrong contacts destabilise the solver immediately on stacks.
+
+**Conclusion: P0.1 was bundled with P1.2.** Both landed together: the
+`begin_refill` primitive replaced `clear_contacts` at the narrowphase
+refill callsite, while P1.2's match-and-merge pass ensures the cached
+impulses follow the right contact slot index across the per-frame
+reshuffle that `reduce_candidates` introduces. The 4-cube atomic-sleep
+test (which collapsed when `begin_refill` was activated without
+matching) now passes with warm-start active.
+
+**Verification (landed).**
+- `ManifoldStore.spec.js`: asserts `j_n` / `j_t1` / `j_t2` survive a
+  `begin_refill` + `set_contact` cycle; `clear_contacts` still wipes
+  them; multi-contact preservation works.
+- `PhysicsSystem.spec.js`: 4-cube atomic-sleep test, 16-cube short
+  window stack, and all dependent integration tests pass with warm
+  start active.
+
+---
+
+### P0.2 — Remove `debugger;` from EPA hot path
+
+**Status**: confirmed at `gjk/expanding_polytope_algorithm.js:136`.
+
+A `debugger;` statement is sitting in the EPA expansion loop. Production
+ship-blocker. Likely a leftover from a prior debug session.
+
+**Fix.** Delete line 136 (and 135 / 137 if they're related guard /
+comment). Confirm by grepping `^\s*debugger` across the package — there
+should be zero hits afterward.
+
+**Verification.** `grep -rn "debugger" src/engine/physics/` returns
+nothing.
+
+**Effort**: minutes.
+
+---
+
+## P1 — Largest accuracy/correctness wins (days)
+
+### P1.1 — Closed-form sphere / box / capsule-vs-triangle
+
+**Status**: LANDED in three sub-commits — P1.1a (sphere), P1.1b
+(box), P1.1c (capsule).
+
+**Problem.** The narrowphase concave dispatch (added in commit 28ef0e894)
+decomposes heightmaps / meshes to triangles and runs per-triangle GJK +
+EPA. `Triangle3D` has a degenerate support along its face-normal axis
+(all 3 vertices project to the same value), so GJK precision is poor —
+hence the skipped settle tests in `narrowphase_concave.spec.js` and the
+documented sink-through-heightmap behaviour. `compute_penetration`
+already works around this for sphere-vs-heightmap by using the
+half-space test on the face plane, but the narrowphase still uses
+per-triangle GJK+EPA.
+
+**Files to add** (suggested layout, paralleling existing
+`sphere_box_contact.js` / `capsule_contacts.js`):
+
+- `engine/physics/narrowphase/sphere_triangle_contact.js` ✅ LANDED —
+  closest-point on triangle (uses
+  `computeTriangleClosestPointToPointBarycentric`), distance to sphere
+  centre, depth = R − dist when overlap. Output stride matches
+  `sphere_box_contact`. Includes upfront degenerate-triangle guard
+  (zero-area triangle returns false) and dist=0 fallback to face
+  normal. 11 unit tests covering face, edge, vertex regions,
+  separation, tangent, singular, degenerate, tilted, indexing.
+- `engine/physics/narrowphase/box_triangle_contact.js` ✅ LANDED —
+  SAT over 13 axes (3 box face normals + 1 triangle normal + 9 edge-edge
+  crosses) with the corrected asymmetric `min(push_pos, push_neg)`
+  per-axis MTV (triangle projection is NOT symmetric around its
+  centroid, unlike box's). Three contact-generation branches:
+  - **Box face winner** → reference = box face, incident = triangle.
+    Project triangle into the face's (u, v) basis, clip with 4 axis-aligned
+    Sutherland-Hodgman passes, recover world contacts on the triangle
+    plane.
+  - **Triangle face winner** → reference = triangle, incident = box
+    face most antiparallel to contact normal. Project the box quad
+    onto the triangle plane (basis = AB normalised + tn × AB), clip
+    with 3 general-half-plane passes against the triangle's 3 edges,
+    recover world contacts on the box face plane.
+  - **Edge-cross winner** → identify the relevant box edge (one of 4
+    parallel edges, selected by sign of `-n` projected onto the two
+    perpendicular box axes) and triangle edge, find closest pair via
+    `line3_closest_points_segment_segment`, emit single contact.
+
+  12 unit tests covering separation, degenerate triangle, box face
+  winner (single + multi-point), triangle face winner, deep
+  penetration, +/-Y face flip, self-consistency reconstruction
+  (`tri_pt = box_pt + depth * n`), rotated box, near-corner case,
+  translated box, output indexing.
+- `engine/physics/narrowphase/capsule_triangle_contact.js` ✅ LANDED —
+  multi-point manifold via primary segment-vs-triangle closest-pair
+  + cap-centre sphere queries. The `segment_triangle_closest` helper
+  takes the minimum of 6 candidates: (case 0) segment crosses
+  triangle plane inside the face → distance 0; (cases 1, 2) each
+  segment endpoint vs `computeTriangleClosestPointToPointBarycentric`;
+  (cases 3, 4, 5) `line3_closest_points_segment_segment` against each
+  triangle edge. Primary contact emitted from the closest pair; each
+  cap centre then runs `sphere_triangle_contact` and is spatially
+  deduplicated against the primary. Result: 1..3 contacts.
+  Pattern matches `capsule_box_multi_contacts`. 10 unit tests.
+
+**Wiring.** Inside `narrowphase_step.js`'s concave branch (the one that
+calls `decompose_to_triangles`), per-triangle dispatch by convex
+shape type:
+
+```js
+if (convex_shape.isUnitSphereShape3D)   → per-triangle sphere_triangle_contact    [LANDED]
+else if (convex_shape.isBoxShape3D)     → per-triangle box_triangle_contact       [LANDED]
+else if (convex_shape.isCapsuleShape3D) → per-triangle capsule_triangle_contact   [LANDED]
+else                                     → existing per-triangle GJK+EPA fallback
+```
+
+Each closed-form path reads the triangle's three vertices from the
+decomposition buffer at stride 10, rotates them to world space (3
+quat rotations per triangle — ~63 flops, negligible vs. the GJK + EPA
+cost being replaced), runs its solver, then runs the same one-sided
+face-normal rejection and normal-based dedup as the EPA path. The
+closed-form paths use ACTUAL surface witnesses (closest point on
+triangle, sphere surface point) rather than body centres — exact in
+both directions, unlike the EPA fallback where support witnesses can
+be arbitrarily far from the contact patch.
+
+**Verification (P1.1 fully landed).**
+- 11 unit tests in `sphere_triangle_contact.spec.js`.
+- 12 unit tests in `box_triangle_contact.spec.js`.
+- 10 unit tests in `capsule_triangle_contact.spec.js`.
+- The two sphere settle tests in `narrowphase_concave.spec.js`
+  un-skipped and passing.
+- The torus-knot test in `PhysicsSystem.spec.js` un-skipped and
+  passing.
+- Full physics suite: 686 passing (was 650 pre-P1.1a — delta:
+  +11 sphere_triangle + 12 box_triangle + 10 capsule_triangle unit
+  tests + 2 un-skipped sphere settle tests + 1 un-skipped torus-knot
+  test).
+
+**Optional follow-up** (not blocking).
+- Tighten `compute_penetration.spec.js`'s mesh-cube test (currently
+  documented as approximate due to closed-mesh over-reporting on side
+  faces). Note: `compute_penetration` is a separate code path from
+  `narrowphase_step.js` — it uses its own GJK+EPA / MPR pipeline and
+  the half-space workaround; landing the narrowphase fast-path
+  doesn't change it. Refactoring `compute_penetration` to share the
+  sphere / box / capsule-triangle fast-paths is a separate cleanup.
+
+**Effort**: LANDED.
+
+---
+
+### P1.2 — Feature-id contact tracking across frames
+
+**Status**: LANDED.
+
+**Problem.** Even after the P0 warm-start wipe is fixed, the manifold
+slot reuses contact-index ordering across frames. Contact slice indices
+shift due to:
+
+- Depth-spread reduction in `reduce_candidates` (`narrowphase_step.js:104`)
+  — picks deepest, then max-spread; ordering depends on candidate
+  generation order which depends on shape-pair details.
+- Box-box face clipping reorders contacts based on which clipped
+  vertices survive (`box_box_manifold.js`).
+- The triangle decomposition emits cells in row-major order — a sphere
+  rolling along a heightmap sees different cell indices each frame.
+
+Reviewers cite two reference designs:
+
+1. **Parry's `TrackedContact { fid1, fid2 }`** — each contact carries
+   feature ids on both sides; the cache matches by `(fid1, fid2)` pair.
+2. **Jolt's `mContactPointPreserveLambdaMaxDistSq`** — match by
+   local-space position proximity (default 0.04 m squared).
+
+We're already laid out for (1): the decomposition pipeline emits stable
+feature ids per triangle, and the convex primitive narrowphase paths
+can derive feature ids from vertex / edge / face indices.
+
+**What landed.**
+
+1. `CONTACT_STRIDE` extended from 13 to 14 (added `feature_id` at
+   offset 13). `CANDIDATE_STRIDE` extended from 10 to 11 in
+   `narrowphase_step.js` (mirroring slot layout).
+2. `set_contact` gained an optional `feature_id` parameter (default 0
+   — preserves backwards-compatible call signatures).
+3. New `clear_impulses(slot, idx)` method on `ManifoldStore`: zeroes
+   only the j_n / j_t1 / j_t2 fields at one contact index. Used by
+   the match-and-merge pass when a new candidate has no matching
+   prev-frame contact.
+4. New `feature_id_of(slot, idx)` accessor for testability.
+5. Per-dispatch feature_id derivation in `narrowphase_step.js`:
+   - `sphere_sphere`: fid = 1 (single contact, real feature)
+   - `sphere_box`: fid = voronoi region 1..27 via new
+     `sphere_box_voronoi_fid` helper (inverse-rotates sphere centre
+     to box-local frame, buckets each component into {<-h, [-h,h], >h})
+   - `box_box`: fid = 0 (clipped-contact ordering isn't cross-frame
+     stable; position-fallback handles per-contact matching)
+   - `capsule_capsule`, `capsule_sphere`: fid = 1
+   - `capsule_box` multi-point: fid = k+1 (per-sub-contact index)
+   - Triangle (concave) path: fid = triangle's pre-existing
+     decomposition fid
+   - GJK+EPA fallback: fid = 0 (no good source)
+6. Match-and-merge pass replaces `clear_contacts(slot)` at the
+   narrowphase refill callsite:
+   - Snapshots prev contacts' `(feature_id, world_a, j_n, j_t1, j_t2)`
+     upfront into scratch (decouples read from write to avoid hazards
+     when the matching mapping shuffles indices)
+   - For each new candidate, finds best prev match — feature_id first
+     (only if both sides have fid ≠ 0), position-fallback within
+     `MATCH_TOL_SQR = 0.02 * 0.02` second
+   - Each prev contact can be claimed by at most one candidate
+     (`prev_claimed[]` flags)
+   - Calls `begin_refill(slot)` (count → 0, data preserved)
+   - Writes each new candidate at slot index k; copies matched prev
+     impulses to slot index k OR calls `clear_impulses(slot, k)` for
+     unmatched candidates
+7. Separation path (cand_count === 0) calls `clear_contacts(slot)` —
+   genuine eviction, drops stale impulses before contact is
+   re-established at a different feature.
+
+**Verification.**
+- `ManifoldStore.spec.js`: 5 new tests covering `set_contact`
+  feature_id round-trip, default fid = 0, impulse-preservation under
+  fid'd writes, `clear_impulses` scoping (one index, geometry +
+  feature_id intact), and per-index targeting.
+- `PhysicsSystem.spec.js`: 4-cube atomic-sleep test passes (was the
+  regression that gated P0.1 alone — confirms warm-start activation
+  is now stable for stacks).
+- Full physics test suite: 650 passing (was 645 pre-P1.2 — the
+  delta is the 5 new ManifoldStore tests; the 4-cube atomic-sleep
+  test which transiently regressed under P0.1-alone is now stable).
+
+**Effort**: landed.
+
+---
+
+## P2 — Performance + precision (day-scale)
+
+### P2.1 — GJK separating-axis cache per manifold slot
+
+**Status**: LANDED (second attempt). The first attempt regressed basic
+settle tests and was reverted; this second pass landed the same
+optimisation via a different mechanism (no try/finally) in 6 inert /
+single-activation steps, each independently verified.
+
+**Mechanism — `gjk_core` + `gjk_with_axis`.** Instead of try/finally,
+gjk's body was extracted into `gjk_core(simplex, A, B, dir)` taking
+the search-direction buffer as a parameter. The existing
+`gjk(simplex, A, B)` is now a thin wrapper that sets the module's
+`scratch_dir` to `(1, 0, 0)` and delegates — semantically identical to
+the old code. The new `gjk_with_axis(simplex, A, B, axis_buf, axis_off)`
+creates a `subarray` view into `axis_buf[axis_off..+2]` and passes it
+to `gjk_core`; in-place writes to the view propagate to the caller's
+buffer automatically, so no try/finally writeback is needed. A guard
+catches zero / NaN / Inf seeds and resets to `(1, 0, 0)`.
+
+**Storage — `ManifoldStore.__slot_axis`.** A `Float64Array(capacity * 3)`
+parallel to the data buffer. Zeroed on slot `acquire` (so a recycled
+slot doesn't leak the previous pair's axis). Grows alongside the
+other slot arrays. Public accessors `slot_axis_buffer` and
+`slot_axis_offset(slot)`.
+
+**Activation — narrowphase.** The outer narrowphase loop computes
+`(gjk_axis_buf, gjk_axis_off)` from the slot and passes them through
+`dispatch_pair`. Both GJK + EPA fallback paths (body-level for
+non-closed-form convex pairs; per-triangle for non-(sphere/box/
+capsule) convex shapes against concave shapes) call `gjk_with_axis`
+with these arguments.
+
+**Sequencing — six steps with independent verification.** Each step
+either had no behaviour change (1, 3, 4) or was a small enough
+activation that a regression would be precisely revertable (5, 6).
+Verified at every step against the full physics suite:
+
+| Step | Change | Tests passing |
+|---|---|---|
+| 1 | Extract `gjk_core` from `gjk`; existing wrapper preserves semantics | 687 / 687 |
+| 2 | Add `gjk_with_axis` + 6 unit tests | 693 / 693 |
+| 3 | Add `ManifoldStore.__slot_axis` (dead code) | 693 / 693 |
+| 4 | Plumb axis args through `dispatch_pair` (dead code) | 693 / 693 |
+| 5 | Activate at the body-level GJK + EPA fallback | 693 / 693 |
+| 6 | Activate at the per-triangle EPA path in concave dispatch | 693 / 693 |
+
+**Why this attempt worked where the first didn't.** The first attempt
+wrapped the existing gjk body in try/finally to write back the cached
+axis on every return path. That should have been semantically
+identical for existing (cache-less) callers, yet basic settle tests
+regressed. Working theory: V8's JIT shape-inference handled the
+try/finally-wrapped function differently in ways that surfaced in the
+existing physics tests' numerical paths. The second attempt avoids
+try/finally entirely — `gjk_core` takes `dir` as a parameter, and
+writeback is automatic via shared memory (the caller's buffer IS the
+working buffer through a subarray view).
+
+**Improvement.** Bullet's `m_cachedSeparatingAxis` and Jolt's `ioV`
+in/out pattern store the last successful separation direction (or
+overlap-resolution direction) per manifold slot. Reuse next frame as
+GJK's initial direction. Reviewers cite 5–10× iteration reduction for
+quiescent persistent contacts — the common case in any settled stack.
+
+**Plan.**
+
+1. Add 3 floats to `SLOT_META_STRIDE` (or a parallel `Float32Array(slot_count * 3)`)
+   in `ManifoldStore` for the cached axis.
+2. Add an overload (or new entry point) to `gjk` accepting an initial
+   direction. Backward-compatible: existing callers (shape_cast,
+   compute_penetration, overlap) pass `null` and default to the
+   current `(1, 0, 0)`.
+3. In `narrowphase_step.js`'s `dispatch_pair`, before the GJK+EPA
+   fallback (line 312 area), read the cached axis from the slot; pass
+   it to GJK. After GJK terminates (with or without overlap), write
+   the final search direction back to the slot.
+
+**Verification.**
+- Microbench: count GJK iterations on the existing
+  `PhysicsSystem.bench.spec.js` torus-knot scene (currently skipped
+  for accuracy but useful here for iteration count). Expect 5–10×
+  reduction.
+- Determinism: same scene state → same iteration counts.
+
+**Effort**: 1 day.
+
+---
+
+### P2.2 — Adaptive EPA tolerance + bail-out fix
+
+**Status**: PARTIALLY LANDED. Adaptive tolerance landed; bail-out
+magnitude consistency reverted (caused energy injection in closed-form
+dispatch paths that share the EPA depth metric — sphere-vs-sphere
+stack regressions launched bodies metres straight up).
+
+**Two issues.**
+
+1. `EPA_TOLERANCE = 0.0001` is absolute. For small-depth contacts (sub-mm)
+   this is comparable to the depth itself; EPA terminates with whatever
+   intermediate face it has, producing a noisy normal direction. Jolt
+   uses a relative tolerance scaled by current closest-face distance.
+2. The bail-out path at line 343 uses `dot_p_search_dir` where the
+   converged path uses `min_dist`. Inconsistent under iteration-cap
+   exit — the documented "best-effort fallback" returns a different
+   magnitude than the converged path of the same geometry.
+
+**Plan and outcome.**
+
+1. ✅ LANDED. Tolerance is relative: `EPA_TOLERANCE_REL = 1e-4`
+   floored by `EPA_TOLERANCE_ABS = 1e-6`. Terminate when
+   `dot_p_search_dir − min_dist < EPS_REL * max(|min_dist|, EPS_ABS)`.
+   This fixes EPA's failure-to-converge for sub-mm contacts where the
+   old absolute tolerance was larger than the depth itself.
+2. ❌ REVERTED. Swapping the convergence-path result magnitude from
+   `dot_p_search_dir` to `min_dist` (for consistency with the
+   iteration-cap bail-out path) was tried — and caused energy
+   injection in closed-form dispatch tests that share the EPA depth
+   metric (sphere-vs-sphere with restitution 0 launched a ball to 7
+   m). The convergence path retains its original magnitude. The
+   "inconsistency with bail-out" the reviewer noted is real but
+   benign in practice: bail-out fires on smooth-shape EPA failure
+   modes that the closed-form paths now bypass entirely (P1.1
+   sphere/box/capsule-vs-triangle); the consistency gap doesn't
+   surface anywhere observable.
+
+**Verification.**
+- EPA spec needs new cases at very small depth (sub-µm) confirming
+  termination is now stable and direction is well-defined.
+- The shape_cast bail-out fallback comparison should produce
+  consistent normals between converged and non-converged exits.
+
+**Effort**: ½ day. Less critical post-P0 because `shape_cast` and
+`compute_penetration` now use MPR for shallow-overlap normal recovery
+— this fix is for the GJK+EPA fallback in `narrowphase_step.js` only.
+
+---
+
+## P3 — Mitigation improvements (half-day each)
+
+### P3.1 — Wire MPR as EPA fallback for mesh-involving pairs
+
+**Status**: LANDED.
+
+`gjk/mpr.js` was already implemented and tested; `shape_cast.js` and
+`compute_penetration.js` already used it. The two narrowphase EPA call
+sites — the per-triangle EPA path inside the concave dispatch (when
+the convex shape is not sphere/box/capsule) and the body-level
+GJK+EPA fallback — now both fall back to MPR when EPA returns
+zero / negative / NaN / Inf depth. MPR's portal-refinement convergence
+properties tend to succeed where EPA's face-expansion hits its
+iteration cap.
+
+**Plan.**
+
+1. In `dispatch_pair`'s GJK+EPA fallback, after EPA returns:
+   - If depth > 0 and direction passes the body-centre sanity check
+     → keep the EPA result.
+   - Else → re-run with MPR (same simplex inputs are not needed; MPR
+     starts from the PosedShape pair). If MPR returns true with a
+     positive depth, use that.
+2. For pairs involving a non-convex shape (heightmap/mesh dispatch via
+   the concave path), once the closed-form triangle solvers from P1.1
+   land, MPR fallback is only needed for the GJK+EPA leg of the
+   concave loop — same pattern.
+
+**Verification.**
+- The torus-knot test in `PhysicsSystem.spec.js` (currently skipped)
+  should make progress with the MPR fallback even before P1.1 lands.
+- `compute_penetration.spec.js`'s smooth-shape direction-precision
+  tests can tighten their tolerances.
+
+**Effort**: ½ day.
+
+---
+
+### P3.2 — Box-box edge-edge closest-pair manifold
+
+**Status**: LANDED.
+
+In `box_box_manifold.js`'s edge-cross-axis-winner branch, the previous
+body-centre-midpoint fallback (both contact points collapsed to the
+midpoint of the two box centres — wrong lever arms, slow drift in
+skewed-box-stack scenarios) is replaced with a proper closest-pair
+computation:
+
+1. Decode `(i, j)` from `best_source = 6 + i*3 + j` — box A's local
+   axis index and box B's.
+2. Pick A's edge among 4 parallel edges along axis `i` by signing the
+   perpendicular axes against `-n` (the direction from A's centre
+   toward B). Same for B with `+n`.
+3. `line3_closest_points_segment_segment(out_st, edge_a, edge_b)`
+   returns `(s, t)`; reconstruct `contact_on_A = edge_a_p1 + s * (edge_a_p2 - edge_a_p1)`
+   and similarly for B.
+
+The "second contact for near-parallel edges" sub-improvement from the
+original plan turned out to be unnecessary in practice: when two box
+edges are near-parallel, the SAT axis `cross(edge_dir_a, edge_dir_b)`
+is degenerate (magnitude below `PARALLEL_EPS_SQR`) and the test_axis
+helper rejects it before it can win. Near-parallel-edge configurations
+end up with a face-axis SAT winner, which the face-clipping path
+already handles with multiple contacts.
+
+**Improvement.** When SAT identifies an edge-edge separating axis
+between two boxes (the cross-product axes in `box_box_manifold.js`),
+the current fallback emits a single contact at the midpoint of the
+edges' projections. The codebase already has
+`core/geom/3d/line/line3_closest_points_segment_segment.js` from a
+previous slice. Use it to emit the proper closest-pair, then evaluate
+spread for a second contact when the edges are near-parallel (within
+some `EDGE_PARALLEL_COS_TOL` like 0.999).
+
+**Plan.**
+
+1. In `box_box_manifold.js`'s edge-edge branch, replace the midpoint
+   contact with `line3_closest_points_segment_segment`'s output.
+2. Detect near-parallel edges via the cos-angle test on the edge
+   direction vectors; in that case emit a second contact at the other
+   endpoint of the parallel overlap interval. Standard SAT-with-clipping
+   pattern.
+
+**Verification.**
+- A box stacked at an angle on another box should reach steady state
+  without the slow drift that the single-midpoint contact causes.
+  Add a regression test alongside the existing box-box stack test.
+
+**Effort**: ½ day.
+
+---
+
+## P4 — Strategic items (weeks; defer until pressed)
+
+### P4.1 — Split-impulse architecture → TGS substepping
+
+**Status**: PLAN.md documents this with the three interacting failure
+modes; reviewers confirmed the path forward in detail.
+
+**Recommendation.** Don't pursue until tall stacks become a frequent
+gameplay pain point. The current 4-cube atomic-sleep test passes; the
+16-cube short-window test catches manifold regressions. The pure-JS
+single-threaded budget is the more likely scaling wall — TGS doesn't
+help there.
+
+If/when we do it: Box2D-Lite-style split impulse — separate
+position-correction pseudo-velocity that doesn't contaminate real
+velocity, restitution as one-shot impulse at first contact (not a bias
+inside the loop), forces apply once at full dt before substeps.
+
+**Effort**: weeks.
+
+### P4.2 — Per-body linear CCD shape-cast
+
+**Status**: PLAN.md backlog ("Per-body linear CCD shape-cast").
+
+**Recommendation.** Already a focused future-work item with the 1km
+falling-tower reproducer (~180/1000 bodies tunnel). Worth scheduling
+once a gameplay use case demands it (e.g. fast projectiles, vehicle
+physics).
+
+**Effort**: week.
+
+---
+
+## Not prioritised in the executive summary
+
+### Shape-cast bisection vs GJK conservative advancement
+
+Reviewers noted Parry's `gjk::directional_distance` (3–5× fewer GJK
+calls per query) but did not prioritise it. Our current bisection +
+slab-narrowing is functionally correct and acceptable; flag for revisit
+if `shape_cast` shows up in profiles. Effort would be a day.
+
+---
+
+## Out of scope
+
+Confirmed from the consolidated review — these continue to be valid
+non-goals:
+
+- SIMD / WASM — V8 doesn't expose deterministic Float64x2 ops.
+- `SharedArrayBuffer` / multi-threaded solver — COOP/COEP headers
+  aren't always available.
+- Cross-runtime bit-exact determinism — would require a soft-float
+  library; same-runtime determinism covers single-device replay +
+  lockstep clients on matched JS engines.
+- Reduced-coordinate articulations — game-physics audience runs in
+  maximal coordinates by convention.
+
+---
+
+## What's already strong (no action; for record)
+
+All three reviewers independently called these out as better than at
+least one reference engine. Useful confirmation that the design bets
+in PLAN.md paid off.
+
+- End-to-end SoA layout with generation-tracked stable IDs and
+  min-heap free list (deterministic ID reuse).
+- Atomic per-island sleep with chain wake via circular doubly-linked
+  `sleep_group_next` / `sleep_group_prev` — wake propagation
+  completes in one frame, vs Bullet/Jolt's bottom-up multi-frame
+  propagation.
+- Deterministic island roots (union-find with union-by-min-index) —
+  Rapier and Jolt do not strictly guarantee this.
+- Flat shape hierarchy, linear narrowphase dispatch,
+  monomorphic-when-hot `support()`.
+- Coulomb-cone disk friction clamp (true cone) — Bullet's default is
+  the anisotropic box clamp.
+- One-sided face-normal rejection on concave dispatch.
+
+---
+
+## Recommended sequencing
+
+1. **P0.2 + P0.1 infra prep (LANDED)** — `debugger;` removed from EPA;
+   `ManifoldStore.begin_refill` added. Surgical, no behaviour change.
+   The initial attempt at full P0.1 activation (flipping the callsite
+   to `begin_refill`) was reverted after the 4-cube atomic-sleep test
+   caught a stack collapse, confirming the reviewer's "slot ordering
+   is usually preserved" assumption is false for stacks.
+2. **P1.2 + P0.1 activation (LANDED)** — feature-id contact tracking
+   and the match-and-merge pass. Warm-start now active and stable on
+   stacks; full physics suite green (650 passing).
+3. **P1.1 in three sub-commits — LANDED**:
+   - **P1.1a (sphere-vs-triangle).** `sphere_triangle_contact` module
+     + spec + narrowphase wiring; un-skipped two settle tests in
+     `narrowphase_concave.spec.js`.
+   - **P1.1b (box-vs-triangle).** `box_triangle_contact` module (SAT
+     + clipping for both face winners + closest-pair for edge-cross)
+     + spec + narrowphase wiring; un-skipped the torus-knot settle
+     test in `PhysicsSystem.spec.js`.
+   - **P1.1c (capsule-vs-triangle).** `capsule_triangle_contact`
+     module (segment-triangle closest-pair + cap-centre sphere
+     queries with spatial dedup) + spec + narrowphase wiring.
+4. **P2.1 (GJK axis cache) + P2.2 (EPA tolerance) together** — both
+   are localised, neither depends on the others.
+5. **P3.1 (MPR fallback) — LANDED.** Wired at both narrowphase EPA
+   call sites (per-triangle concave + body-level fallback). Behind
+   the existing depth-validity gate so it only fires when EPA fails.
+6. **P3.2 (box-box edge-edge) — LANDED.** `line3_closest_points_segment_segment`-based
+   contact-pair reconstruction; near-parallel edges already handled
+   by SAT degeneracy fallthrough to face axes.
+7. **P4 items**: defer until gameplay pressure.