deps: bump the splat-renderer fork to upstream 8.0.1

[annejan]

Repoints [patch.crates-io] at the §1–11 edit set replayed onto upstream 8.0.1 (f94855a4 → ffaecd70), and follows the version requirement in Cargo.toml. No martin source changes; no shader behaviour change.

What comes in

Two upstream commits:

• #236 — the 8.0.1 version bump.
• #237 — the lazy_type_alias → checked_type_aliases nightly feature-gate rename, which is exactly the fix we were carrying as fork edit §11. That edit is now a docs-only entry instead of a code divergence, so the fork is one edit closer to upstream.

Also folded in: upstream's chunks_exact → as_chunks modernisation in query/raycast.rs and its test/tool code.

The rebase was verified, not trusted

The rebased head already existed on the fork, pushed by an automated session a couple of hours earlier, so it was re-derived rather than taken at face value:

• Replaying martin-tightcut onto upstream/main locally reproduces the pinned commit's tree byte-for-byte.
• Diffing that commit against upstream leaves only our seven edited files — CHANGES.md, gaussian/settings.rs, io/scene.rs, morph/interpolate.wgsl, render/bindings.wgsl, render/gaussian.wgsl, render/mod.rs — with all §1–11 edits intact and zero conflicts.

⚠️ Follow-up needed on the fork repo

The pinned rev currently lives on the fork's martin-tightcut-rebased branch; martin-tightcut still points at the 8.0.0 base. This session had no push access to annejan/bevy_gaussian_splatting, so the canonical branch wants a fast-forward:

git push --force-with-lease origin ffaecd70f5102934beaca58845d3d2299b73394a:martin-tightcut

Do not delete martin-tightcut-rebased before that — the pinned rev would become unfetchable and fresh checkouts would break. A note to this effect sits next to the pin in Cargo.toml.

The fork's own CHANGES.md is also now stale in two places: the header still says "upstream 8.0.0", and §11 still ends "Upstream (mosure) will need the same fix" — it has since landed as #237.

Nothing else was behind

cargo update moves zero packages. Every direct dependency is already at its latest published version — Bevy 0.19.1, wgpu 29.0.4, resvg 0.48.1, image 0.25.10, clap 4.6.6, fundsp 0.23.0, rayon 1.12.0, serde_json 1.0.151. The only crate in the tree behind a newer release is generic-array 0.14.7, held there transitively (1.x is a breaking rewrite). The SHA-pinned Actions are Dependabot-managed (last grouped bump 2026-08-28, #35), and .cargo/audit.toml still needs only the one ttf-parser unmaintained-crate ignore.

The toolchain pin stays at nightly-2026-08-28, deliberately. Newer nightlies exist, but that pin is two days old, was whole-tree verified, and carries the measured -Znext-solver=coherence opt-out; re-rolling it for two days of churn buys nothing and risks the compo build.

Verification

Every CI gate, run locally and green:

Gate	Result
cargo fmt --all --check	pass
cargo clippy --all-targets -- -D warnings	pass
cargo test --release	176 passed, 0 failed
reuse lint	compliant (253/253 files)
sh3 profile (--no-default-features --features sh3)	checks clean
strict score validate (MARTIN_SCORE_STRICT=1)	pass
intro --features bundle single binary	builds (4 assets, 5288 KiB → 5016 KiB)

Plus a real headless render through software Vulkan (lavapipe), which is stronger than CI's own smoke job — that one usually goes inconclusive on the free runners:

• probe frame at t=2 renders the deFEEST logo correctly, with bloom.
• pipeline/smoke-shows.py intro — all three probes non-black (mean=40.4 / 49.9 / 6.2, lit=0.397 / 0.537 / 0.047).

---

Generated by Claude Code