hpc::phase: phases in turns, a (cos, sin) LUT with stated error bounds, a drift-free accumulator - #343

This PR adds src/hpc/phase.rs, re-exported from ndarray::simd. It is the "ADOPT" item of the lance-graph phasor probe (D-PHT-1, lance-graph #1416).

What

Phases. A phase is a u32 in turns, with 2^32 = one rotation. So 3θ is p.wrapping_mul(3), θ + φ is p.wrapping_add(q), and wraparound is free.

PhaseLut methods:

• nearest(p) and lerp(p) (linear interpolation), plus lerp_batch.
• Per-component error bounds, so a caller can carry them into an exact decision:
  • nearest_error_bound() is π/2^bits plus f32 rounding;
  • lerp_error_bound() is h²/8 (chord sag, with h = 2π/2^bits) plus 4·f32 epsilon.

Why

ndarray had no phase LUT, no CORDIC and no SIMD sin/cos. vml::vscos/vssin call scalar cos/sin per lane.

The lance-graph probe measured interpolation at 2^12 entries:

• Accuracy: the same as f32 sin_cos (4.6e-7 vs 5.1e-7 relative).
• Speed: about 4.5× faster than f32 sin_cos and 12× faster than f64 (x86-64-v3, release, 1M random phases).
• Drift: an integer accumulator has none; an f32 complex recurrence reached 1.95 units on R = 100 after 1M steps.

The lookups here are scalar. No SIMD path is claimed. A gather batch is a separate, unmeasured step.

Tests

• Bounds, checked two-sided:
  • lookups never exceed their bound, for 4, 8, 10, 12 and 16 bits;
  • the bounds are not vacuous: nearest reaches more than half of its bound, and lerp more than a tenth at ≤ 12 bits;
  • interpolation beats the nearest entry.
• Wrap and multiply: the last interval interpolates into entry 0, and wrapping_mul(3) equals 3θ mod 2π.
• Conversions: they round-trip, negative angles wrap, NaN maps to 0, and from_radians reports the step it actually took.
• Accumulator: after 1M steps the phase equals the closed form, and the looked-up point stays within the bound.
• Batch and input checks: the batch equals the scalar lookup; mismatched lengths and out-of-range sizes panic.
• Doctests: 20, one per public item.

Disable runs, each red then restored:

• nearest without the half-step rounding;
• lerp without the wrap into entry 0;
• the lerp bound without the chord sag;
• at ignoring start.

Gates:

• cargo test -p ndarray --lib: 2543 passed.
• cargo clippy -p ndarray --lib --tests -- -D warnings: clean.
• fmt: clean.
• Every compile used debug 0. Only the native realization was run; the module has no target-specific code.

The blackboard entry is prepended.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EFw2WdKr1oxvaKCJC2ua2R

Generated by Claude Code