fix(gjk): shape-cast TOI error scales with target shape extent

[vikng-dev]

Fixes #429.

What

cast_shapes / minkowski_ray_cast returns a time of impact that comes back short by an amount that scales with the target shape's extent. Casting a ball (r = 0.5166) straight down at a cuboid, measured max TOI distance error:

cuboid half-extent	before	after (this PR)
5 m	0.246 mm	0.113 mm
50 m	3.64 mm	0.0023 mm
500 m	139.45 mm	0.00024 mm

The witness data (point1/normal1) is exact even when the TOI is wrong.

Reproduction

git checkout 89eb49a   # test commit only, no fix
cargo test -p parry3d --test lib shape_cast_toi_accuracy
#   FAILS — prints the error table above (139 mm at 500 m half-extent)
git checkout fix/shape-cast-stagnation-refine
cargo test -p parry3d --test lib shape_cast_toi_accuracy   # passes

Mechanism

At first I suspected the relative convergence tolerance (gjk.rs:775-786), but it never fires for this case — instrumented over the test's 2,400 casts: 0 hits. What actually happens: minkowski_ray_cast advances a lower TOI bound ltoi and tracks an upper bound. With large support coordinates orthogonal to the cast direction, float cancellation keeps the upper bound from decreasing; the "upper bounds inconsistencies" last-chance exit fires (gjk.rs:712-715) and returns the current ltoi as the hit (gjk.rs:721) — still short by the stagnated gap. All 2,278 erroneous TOIs in the instrumented run exited through this path.

Fix

On the last-chance path, refine ltoi from the simplex witnesses already in hand before accepting the impact: project the witness separation along the cast direction ((witness1 - witness2 - ray.origin) · dir). The witnesses stay precise exactly where the upper bound stagnates, so this recovers the accuracy the bound lost. Witness and normal generation are untouched — the returned contact data is bit-identical to before; only the TOI scalar changes. Added cost is confined to the rare stagnation exit (≤ dim+1 weighted witness accumulations via the existing result()).

Two intentional behavior changes to be aware of:

1. The refined TOI is no longer a certified lower bound: measured overshoot up to +0.113 mm at 5 m half-extent (curvature converts lateral witness deviation into forward error ≈ r·θ²/2), where the old exit was strictly short. Three orders of magnitude below the old error, but callers relying on "TOI never long" should know.
2. A stagnation hit whose refined TOI exceeds max_time_of_impact now returns None, where the old code returned a wrongly-short hit inside the limit.

Real-world impact

Found in a 64 Hz networked tank sim: wheel sphere-casts against a 1000 m ground cuboid came back short by p50 33 mm / p99 134 mm / max 200 mm over 19,828 samples; through a 551 kN/m suspension spring that was 10-40 kN of per-tick force noise — a sustained at-rest hull limit cycle (~12 mm heave) and a standing client/server divergence amplifier. #414 hit the same exit independently from a Rapier DynamicShapeCastVehicleController (wheels vs a 10,000-unit ground backstop, TOI ~10× short) — this PR is the same diagnosis with a regression test and the refine kept inside the existing exit. Possibly related: #180's pose-knife-edge None returns come from this same function's exit structure, though this PR does not address that case.

Tests

shape_cast_toi_accuracy_does_not_scale_with_shape_extent (tests/geometry/time_of_impact3.rs) pins max error ≤ 0.2 mm across half-extents 5/50/500 m and exact witnesses. Full workspace --tests suites pass for parry3d, parry2d, parry3d-f64, parry2d-f64; cargo fmt --check clean.

[Ughuuu]

In the test you added, before and after what is the error? And/or at extreme distances, what is the error? (I see in example you put before it was 139mm, and after? 0mm or?)

And this seems to change the gjk algorithm (for the case of error? So it doesn't converge enough or?), is this the only source of error? (eg. there is also a path with epa for eg. in some cases as I remember).

EDIT:

Oh, ok, so after your PR 0.00024 mm error, before 139mm error.

[vikng-dev]

Aye, 139mm -> 0.00024mm.
GJK iteration is unchanged, the change only touches the existing "upper bounds inconsistencies" (gjk.rs:712). The problem isn't convergence, in my tests the convergence tolerance never fired - all the wrong TOIs came from the stagnation exit.
epa isn't on this path, cast_shapes is pure gjk. cast that starts in penetration returns TOI = 0 so it's never reached.