feat(examples): stress-tv autosweep measures load time, not steady-state FPS

[chiefcll]

Why

On a 60 Hz TV, the steady-state FPS of a static grid just pins at the panel refresh — so the old autosweep reported every tier as "limited by index cap" at ~60 fps. The FPS number revealed nothing, because the renderer keeps up frame-to-frame right until the correctness cap. The cost that actually hurts a constrained device is the one-time build + texture upload (the slow load, eventual OOM), which the FPS sweep deliberately skipped during its warm-up.

What changed

• The sweep now measures time-to-ready (buildGrid → renderer idle) per count and reports the highest count that loads within a budget:

  ?test=stress-tv&autosweep=true                (default 1000ms)
  ?test=stress-tv&autosweep=true&loadbudget=500
  

• FPS kept as a secondary readout (labelled, not the deciding metric).
• Stops probing past the ladder max. The old code appended the index cap (up to 16384) as an extra test rung — fine for cheap FPS sampling, but building 16384 nodes for a load-time measurement would hang/OOM the very device we're profiling.
• Per-tier limiter is now meaningful: ladder max / >Nms load / index cap, instead of always index cap.

Verified

In-browser (tight loadbudget=80 to force a crossover on a fast 120 Hz dev box):

SWEET SPOT - loads in <80ms   (VAO ON)
tier 1: 2000 cards  (42ms load, 120 fps, ladder max)
tier 2: 1775 cards  (77ms load, 120 fps, >80ms load)   ← bisected
tier 3: 1200 cards  (72ms load, 120 fps, >80ms load)
tier 4:  800 cards  (50ms load, 120 fps, index cap)

Each tier now reports a distinct, meaningful limit. Default scene (200 cards, +image) is unchanged, so the certified chromium-ci snapshot is unaffected.

🤖 Generated with Claude Code