Physics3D: contact-constraint overflow at ~10k bodies makes objects tunnel through static geometry; JoltScene ignores PhysicsSettings limits #523

Is your feature request related to a problem? Please describe.
At ~10 000 contacting dynamic bodies, Jolt's contact-constraint buffer overflows and bodies silently tunnel through static geometry. The perf stress battery's physics_pile_10000 scene (10k boxes/spheres falling into a walled arena) spams

[error] OloEngine: Jolt physics update error: 4

(4 = EPhysicsUpdateError::ContactConstraintsFull — "Some contacts were ignored. Increase inMaxContactConstraints") and the battery's workload probe caught a body at y = −36 and still falling after 20 s — through the arena's static floor (its collider was valid; the 1k variant of the same scene settles correctly at y ≈ 0.5).

Two config problems compound it:

1. JoltScene.h hardcodes s_MaxContactConstraints = 10240 (and s_MaxBodies/s_MaxBodyPairs = 65536) as constexpr, so a project cannot raise it.
2. PhysicsSettings::m_MaxContactConstraints (PhysicsSettings.h:28) exists and is honoured by Physics3DSystem::Init — but JoltScene::Init (JoltScene.cpp:2383) uses its own constants and ignores the settings. The two init paths disagree.

Describe the solution you'd like

• Route JoltScene init through PhysicsSettings (one source of truth), with defaults sized for the advertised body counts (10240 constraints is far too small for 65536 max bodies).
• Surface the overflow loudly but throttled (see the FrameDataBuffer log-spam issue): today it's an every-step error line, and the gameplay symptom (objects falling through the world) is disconnected from the cause.

Describe alternatives you've considered

• Documenting a practical body-count ceiling instead — but 10k dynamic bodies is within Jolt's comfortable range when the buffers are sized for it; the engine shouldn't corrupt the simulation silently.

Additional context
From the perf stress battery (generate_perf_scenes.py --scene physics_pile --count 10000, Dist). At 50k bodies MaxBodies (65536) still fits, but the same overflow applies. Frame cost at 10k was ~1.6 s/frame in Dist (fixed-timestep catch-up spiral) — separate scalability data in the battery findings report.

Score

capability: 3
craft: 3
stability: 5
decay: 1
effort: 2
confidence: 0.9
learning: 2
fun: 3
kano: table-stakes
blocked_by: []
blocks: []