perf: pack reference orbit to 16 bytes/point, fast-extend periodic orbits, and scale deep-zoom slice ceiling

[rowan-m]

Summary

• Halve reference_orbit buffer size (32 bytes/point -> 16 bytes/point): Store each reference orbit point as a single vec4<f32> ([re_hi, re_lo, im_hi, im_lo]) using split_f64_to_2_f32 instead of 2 × vec4<f32>. Since reference points are converted from f64 (53-bit significand), two f32s already capture the full Double-Single significand while halving WASM memory allocation, worker transfer size, writeBuffer upload bandwidth, GPU storage cache footprint, and raw_xm register pressure in fs_main_qs.
• Fast periodic orbit extension & deep-zoom periodicity:
  • In wasm/src/lib.rs (Orbit::advance), record the detected cycle period from CycleDetector and, once a full period of p consecutive points is stationary in qs, copy the p-point cycle directly up to limit instead of re-iterating in arbitrary-precision FBig.
  • In src/renderer/shader.wgsl, extend fs_main_f32 Brent periodicity detection across its full zoom range by comparing perturbations delta whenever next_m == period_m (where X_m cancels out), and use fma in ds_mul.
• Scale adaptive slice ceiling at ultra-high iteration counts: In Renderer._sliceRows(), floor maxOps at MIN_SLICE_ROWS * 2 * rowOps so fast exterior slices at ultra-high iteration counts (where 16 * budget < 32 rows) can still scale up to 2 × MIN_SLICE_ROWS (64 rows) while keeping slow interior slices floored at 32 rows.

[github-actions]

Visit the preview URL for this PR (updated for commit fa42de4):

https://fractious-deep--pr98-perf-orbit-packing-a-f1zzexlj.web.app

_{(expires Mon, 05 Oct 2026 15:21:47 GMT)}

_{🔥 via Firebase Hosting GitHub Action 🌎}

_{Sign: 348393a9746312ebab7a45ce5eba16e6d119af34}