perf(sdf): translation fast path for scrolling text + skip unchanged SDF uploads

[chiefcll]

What

Two SDF text hot-path fixes targeting the standard TV scroll scenario (rows of cards with titles translating under a clipped viewport). Follows #119/#120 from the same render-loop review.

1. Translation fast path for the SDF vertex cache

The SDF vertex cache stores world-space vertices, so a scrolling text node missed it on every frame — the hit test required all six transform components to match. Each miss re-ran per-glyph matrix math (16 mul + 16 add per glyph) in addSdfQuads and re-snapshotted the result back into the cache (a second full copy, plus a realloc when glyph count changed) — every node, every scroll frame.

SdfTextRenderer.renderQuads now dispatches three ways:

• Exact hit (unchanged): single mem-copy via addSdfCachedQuads.
• Translation hit (new): scale/rotation components match, only tx/ty moved. The new WebGlRenderer.addSdfTranslatedQuads mem-copies the cached vertices and adds (dx, dy) to the two position floats of each vertex. The cache keeps its original base transform, so every frame recomputes from the same reference — no accumulation drift and no per-frame re-snapshot.
• Miss (unchanged): full recompute + snapshot.

⚠️ Correctness subtlety for reviewers: the copy must remain a typed-array set() (bit-exact memcpy). Packed ABGR colors live in the same Float32Array, and patterns like 0xFFFFFFFF — plain white text — are float32 NaNs that element-wise float reads/writes can canonicalize into 0x7FC00000, corrupting colors. Only the two position floats are touched after the copy, and a regression test asserts bit-exact color preservation through the translated path.

2. Skip the per-frame SDF GPU upload when bytes are unchanged

render() uploaded the entire SDF buffer via bufferData(DYNAMIC_DRAW) on every drawn frame — so any focus pulse or image fade re-uploaded every glyph on screen (96 bytes/glyph; exactly the "guaranteed CPU tax" in the CLAUDE.md TV cost model). This implements the sdfBufferChanged invariant that CLAUDE.md already documents:

uploadSdfBuffer() skips the upload when every write this frame was an exact cache-hit mem-copy (sdfBufferChanged === false) and the total size matches the last upload. Exact hits write byte-identical data, and identical offsets are guaranteed because every source of reorder or resize raises the flag, failing conservative per the project rule:

• cache-miss recompute (addSdfQuads)
• the new translated copy (addSdfTranslatedQuads)
• render-list rebuild (invalidateQuadBuffer) — a reorder moves identical bytes to different offsets
• RTT partial upload (renderRTT) — clobbers the GL buffer's contents/size mid-frame
• SDF backing-store growth (ensureSdfBufferCapacity)

Impact

• Scroll frames: per text node drops from full per-glyph transform + snapshot copy-back to one memcpy + 2 adds per vertex.
• Non-text animation frames (focus pulses, fades — constant in TV UIs): the SDF bufferData driver copy is skipped entirely.
• Scrolling text still uploads every frame (bytes genuinely change) — that's correct; the CPU to produce those bytes is what shrank.

Testing

• Full unit suite: 310 tests pass (20 new)
  • WebGlRenderer.sdfBuffer.test.ts: translated-copy math (positions shifted, UVs/distanceRange untouched), bit-exact NaN-pattern color preservation, append/zero-glyph behavior, upload-skip semantics (skip on identical frame; re-upload on flag or size change), and every flag raiser incl. render-list rebuild and buffer growth
  • SdfTextRenderer.test.ts: three-way dispatch — miss snapshots the cache, exact hit mem-copies, translation hit passes the right delta and does NOT re-snapshot, scale/rotation/color/alpha changes fall back to the miss path
• Manual, examples dev server:
  • stress-single-level-text (100 white SDF texts animating x/y — pure translate path with the NaN color pattern): crisp, correctly colored across frames
  • text, text-ssdf (static), text-scaling (miss path): render correctly, console clean
• pnpm build clean; Prettier/ESLint clean (0 errors)

🤖 Generated with Claude Code