perf(render,wasm): add Order-8 Series Approximation and factor perturbation loops

[rowan-m]

Summary

Maximises deep-zoom CPU/WASM and WebGPU rendering throughput while keeping progressive slices responsive on mobile GPUs:

1. Order-8 Circle-Probed Series Approximation (wasm/src/lib.rs, src/renderer/shader.wgsl):

   • Evolves an 8-term Taylor polynomial $\Delta_m(u) \approx \sum_{k=1}^8 a_k^{(m)} u^k$ normalized to the viewport bounding disk $u = \Delta c / R_{\text{max}}$ ($|u| \le 1$, preventing f64 exponent overflow/underflow at arbitrary zoom depths).
   • Validates the polynomial up to the first close approach / rebase against 8 boundary probes on $|u| = 1$ (by the Maximum Modulus Principle, maximum truncation and non-linearity error over the disk occurs on its boundary).
   • Transfers the Double-Single normalized coefficients to the GPU so pixels inside the validated disk (dot(u_norm, u_norm) <= 1.0) jump directly to i = skip_iter, m = skip_iter (skipping 12,584 iterations per pixel on z = 17.345), while out-of-disk pixels during large interactive drags seamlessly start at 0.

2. Factored Perturbation Recurrence Across All 3 Shader Tiers (src/renderer/shader.wgsl):

   • Replaces $2 X_m \Delta_n + \Delta_n^2 + \Delta c$ with $(2 X_m + \Delta_n)\Delta_n + \Delta c$, eliminating qs_sqr, qc_sq, ds_sqr, and dc_sq (~43% fewer multiplications per iteration with identical floating-point accuracy).

3. Precision Tier Threshold & SA-Aware Slice Budgeting (src/Renderer.js):

   • Extends Tier 2 (DS, Double-Single) to z < 21.0 (scale > 1.0e-21), since post-rebase squaring $Z_n^2 + \Delta c$ only requires $z/2$ digits of mantissa precision.
   • Discounts skipIter in _effectiveIter(config) for interactive preview and progressive slice budgeting, and updates opsMultiplier (12.0 for DS, 8.0 for QS) to reflect the factored shaders and avoid single-wavefront 1-row slice traps on tile-based mobile GPUs.

4. Zero-Allocation f64 Quad Splitting & Early Phase-1 Spiral Exit (wasm/src/lib.rs):

   • Replaces split_fbig_to_4_f32 (12 heap allocations per step) and FBig escape addition in Orbit::advance with pure stack f64 splitting (split_f64_to_4_f32).
   • Exits the Phase 1 3 × 3 FBig spiral in search_anchor as soon as a candidate survives past the immediate exterior (score > 16) so Phase 2 surveys the rest of the viewport in fast hardware f64 perturbation.

[github-actions]

Visit the preview URL for this PR (updated for commit 602986a):

https://fractious-deep--pr96-perf-series-approx-a-ow4olipy.web.app

_{(expires Fri, 02 Oct 2026 15:35:33 GMT)}

_{🔥 via Firebase Hosting GitHub Action 🌎}

_{Sign: 348393a9746312ebab7a45ce5eba16e6d119af34}