Bump bevy 0.18.1 → 0.19: upstream removes the near_clip_plane footgun (rl#196 class)

[bddap-bot]

Bevy 0.19.0 (released) removes the near/near_clip_plane footgun that caused #196: adjust_perspective_matrix_for_clip_plane treats any straight-ahead plane (0,0,-1,*) as non-oblique regardless of distance, so a custom near can no longer be silently overridden by the stale default plane.

On the bump, prune together (noted in fp_perspective(), net/src/render/scene.rs):

• the near_clip_plane: Vec4::new(0.0, 0.0, -1.0, -near) line (dead weight on 0.19)
• the "Bevy 0.18" version-specific doc comments there and on the regression test

The regression test fp_camera_effective_clip_is_the_scaled_near passes on both versions, so it rides through the bump as the guard.

[bddap-bot]

Migration done and green; blocked on a rapier prerelease. Root cause now pinned exactly.

bevy 0.18.1→0.19 is fully implemented on branch bevy-0.19-wip — builds clean, 199/200 tests pass: bevy_egui 0.39→0.41, bevy_rapier3d 0.34→0.35, naga held at 26 + termcolor (burn's cubecl-wgpu still pulls naga 26 and bevy no longer enables its termcolor), shell.nix nixpkgs bumped for rustc ≥1.95. The near_clip_plane line is pruned per this issue and the guard fp_camera_effective_clip_is_the_scaled_near passes on 0.19 — the footgun fix is confirmed.

Blocker. bevy_rapier3d 0.35 exact-pins rapier3d =0.33.0-alpha. That prerelease is where rapier migrated its math types from nalgebra to glam, and bevy_rapier3d 0.35 is written against it (rapier::math::Vec3 is bevy::Vec3). Every released rapier — 0.33.0 (Jun 5) and 0.34.0 (Jul 4) — is still nalgebra-based. I shimmed both in and bevy_rapier3d 0.35 fails identically: 261 errors, all the trait bound bevy::Vec3: From<rapier::math::Vec3> is not satisfied. So no released rapier can satisfy bevy_rapier3d 0.35 — the only usable rapier is that alpha.

And that alpha diverges to NaN in the multibody joint-motor velocity constraint under in-range actuator input (unit_multibody_joint.rs:113 → clamp(NaN,NaN) panic inside step_simulation; curr_pos already NaN). Repro: existing test boarded_ship_spawns_clear_of_a_flailing_sally — the crab's motorised legs hit it in ~200 steps, and the trainer's policy drives the same [-1,1] range, so this would NaN the live trainer, not just the test. That's the 1 failing test, and why this can't land.

Unblocks when rapier ships a stable glam-based release (and bevy_rapier3d pins it), or the alpha's multibody-motor divergence is fixed upstream. Then: rebuild bevy-0.19-wip, re-run the flail test, run the build matrix, land. Not forking rapier to revert upstream multibody math on the trainer's physics. This job re-checks daily.