Address inverse-FFT / edge-detection review findings

[jacobson30-bot]

Fresh review pass over the inverse-FFT and edge-detection modules.

inverse_fft.py (minor)

• Moved from dataclasses import replace out of the per-region loop in fourier_region_mask to module scope.
• Typed FourierStrokes.stamps as tuple[tuple[float, float], ...].

(Even-N Nyquist conjugate limitation and the O(stamps × pixels) stroke cost were reviewed and left as-is — the former is documented and surfaced via imag_residual_norm, the latter is fine for typical brush counts.)

edge_detection.py — gradient_filter

• Medium — no more fabricated edges at NaN borders. Mean-filling non-finite pixels steps the hole rim, so every pixel whose 3×3 operator footprint touches a NaN had a spurious gradient (~15× baseline) while only the NaN pixels themselves were excluded. Now the NaN mask is dilated by the 3×3 footprint and that band is nulled (magnitude→0, orientation→NaN, display→NaN) before normalisation/thresholding, and excluded from the edge mask.
• Low — orientation sentinel. ROI/invalid regions now use NaN for orientation instead of 0.0 (a valid +x direction that conflated "no data" with "horizontal").
• Low — convention documented. gradient_orientation is in image coordinates (row axis points down), noted on both the function and EdgeDetectionResult.

Verification

• 1971 passed, 207 skipped, ruff check clean.
• New regression tests: NaN-border exclusion, clean-image no-op, NaN orientation sentinel outside ROI.

🤖 Generated with Claude Code