perf: reuse Three.js scratch objects in the replay render loop

[Anexus5919]

📌 Related Issue

Fixes #879

---

📝 Description

Replay3DModel's render loop allocated a fresh batch of Three.js objects every frame:

• updateStressVectors cloned several vectors and built two THREE.Colors per joint.
• buildSegmentScaleState round-tripped a Matrix4/Vector3/Quaternion to recover a scale it already knew.
• updateSegmentScaleAdaptor allocated identity vectors as lerp targets.

Across many joints and segments at 60fps this churned the garbage collector (frame hitches).

🔹 What has been changed?

• src/components/Replay3DModel.tsx:
  • Hoisted the constant axes/colours and reusable scratch Vector3/Color to module scope.
  • updateStressVectors: write directly into mesh.position, use the constant axis/colour objects, and reuse the stored previous-position vectors via .copy().
  • buildSegmentScaleState: set the scale components directly into a scratch vector (dropping the unused Matrix4 round-trip; only .scale was ever read).
  • Segment lerp targets use a shared identity vector.

🔹 Why are these changes needed?

• A 60fps render loop should not allocate dozens of short-lived objects per frame. Reusing scratch objects and writing into existing targets keeps per-frame allocations near zero, removing the GC pressure. Each scratch is fully written before use and not retained (the previous-position vectors are reused in place), so behaviour is unchanged.

---

🛠️ Type of Change

• ⚡ Performance Improvement

---

🧪 Testing

✅ Tests Performed

• Tested locally
• npx tsc --noEmit: Replay3DModel.tsx clean (the only pre-existing tsc errors are in exerciseEngine.ts, unrelated).
• npx eslint src/components/Replay3DModel.tsx: 0 problems.
• Reviewed the vector math for aliasing: each scratch is copied/written before use and not stored, and direction is copied into a separate scratch before being scaled, so it remains valid for the quaternion call.

🔹 Note on automated tests

No runtime unit test: Replay3DModel is a Three.js/WebGL component that does not run under the vitest/jsdom environment. Verification is by tsc, eslint, and the aliasing review above (allocation reduction is observable in the DevTools memory timeline).

🌐 Browsers Tested

Not applicable (allocation profile).

---

📷 Screenshots / Demo (if applicable)

Not applicable.

---

📋 Checklist

• I have read the project's CONTRIBUTING guidelines
• My code follows the project style guidelines
• I have performed a self-review of my code
• I have tested my changes locally
• I have added/updated documentation where necessary (not applicable)
• My changes do not introduce new warnings or errors
• This PR is linked to an existing issue

---

💬 Additional Notes

Branched from latest upstream/main (5464425). This touches Replay3DModel.tsx, as do a few other replay PRs (#906, #907, #878); they change different lines and are independently mergeable, with at most a trivial rebase for whichever lands later.

[vercel]

@Anexus5919 is attempting to deploy a commit to the somiljain2024-4175's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Anexus5919]

@Somil450 @diksha78dev Kindly have a review on this pr. Thanks!