fix: stabilize LookAt for parallel up vectors

[besmpl]

What changed

• normalize the caller's up vector before testing LookAt basis stability
• provide deterministic fallback axes for parallel and anti-parallel view/up directions
• preserve conventional +Y orientation for a zero up vector and a valid -Z view for coincident eye/target positions
• add matrix- and node-level regressions for parallel, anti-parallel, near-parallel, scaled, zero, tiny, and ordinary inputs

Why

Mat4LookAt previously normalized forward × up without handling a zero cross product. A camera directly above or below its target with the conventional {0,1,0} up vector therefore produced a singular view matrix. This includes the top-down camera arrangement shown in the project documentation; projected points could end up with an invalid clip-space w.

The fix constructs a stable, right-handed orthonormal basis while keeping ordinary non-degenerate views unchanged.

Impact

Top-down and bottom-up cameras now orient correctly instead of producing a zero basis. Up-vector magnitude no longer affects the result, and boundary inputs remain finite and deterministic.

Verification

• focused Mat4LookAt, Node.LookAt, and camera tests (including repeated and race runs)
• go test ./...
• go test -race ./...
• go build ./...
• go vet ./...
• CGO-disabled Linux, Windows, and Darwin cross-builds
• gofmt and git diff --check
• baseline discrimination and independent review

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[kolkov]

Code reviewed: robust LookAt degenerate case handling — zero direction, zero up, parallel up/forward fallbacks. Bitwise epsilon comparison correct. 10 tests with orthonormal basis validation. LGTM.