[PlasmaDev5]

PH-F36: headless physics regressions and soak gate

Extends the existing JoltP0Tests target into JoltTests, one executable that links the real plJoltPlugin and drives it through plWorld. It needs no editor, no GPU device and no package store.

Groups (127 checks)

p0, stacking, layers, triggers, queries (every query type), breaks, buoyancy (30/60/144 fps), character (steps, ramps, platform, teleport), lifecycle. --stress <s> runs a soak scene (body and joint churn, ragdoll components, walking character, 200 m overlap/raycast sweep).

CI

Tests/Run-JoltTests.ps1 plus .github/workflows/jolt-tests.yml (self-hosted Windows runner, PLASMA_SDK_ROOT repo variable). The workflow itself has not run on a runner yet.

Results (local, main checkout as SDK)

• Development: 127 checks, 0 failures, 6 open R1 items
• Debug: 127 checks, 0 failures, 7 open R1 items (the fixed-constraint break threshold depends on a transient, so it differs between builds)
• Debug soak: 120 s, ~70k frames, no assert or crash. The 10 minute run has not been done.

Open R1 items (checks written against intended behaviour; only --strict fails on them)

• GetGravity() returns a hard-coded -10
• A body in two sub-shapes of one trigger is reported more than once
• m_bIgnoreInitialOverlap sweeps still report the shape they start in
• Fixed-constraint break force uses ReduceMax() of the solver lambda (negative under load) and is a per-step impulse, not newtons

RaycastAll also reports each convex shape twice (entry and exit); tests count distinct actors.

Not covered (needs assets)

Ragdoll with real skeletons, glass slabs under fire, contact-damage destructibles, the AI voxelizer itself. See Tests/README.md.