Fix real Phase 1-4 perf hotspots found by profiling (~5x on a real session)

[hd152]

Summary

Profiling a full pipeline run (not guessing) found several real hotspots across master calibration, DBE, post-processing and Gaussian blur (~30 call sites project-wide). Fixed in stages, each measured before moving to the next — same real profiled session throughout: 3m 7.8s -> 36.9s (~5x).

• robust_pca_decompose fusion: two new native kernels (robust_pca_pre_svd_input, robust_pca_iterate) fuse the per-iteration IALM elementwise arithmetic that was outside the already-native SVD step; the L-update matmul is rerouted through the existing small_times_wide kernel instead of np.matmul (this machine's numpy has no optimized BLAS). 126.9s -> 45.4s.
• Calibration timing visibility: _build_masters's timer was gated on real bias/dark/flat frames existing, so --flat-from-lights (which runs precisely when they don't) was invisible, hidden inside an unlabeled "Other (I/O)" bucket. Now timed unconditionally with its own line.
• --flat-from-lights CFA-aware downsample: block-averages each of the mosaic's 4 Bayer sub-planes independently (never across colours) before decomposition, upsampling the recovered flat back to full resolution after. 45.2s -> 3.6s.
• DBE compact-source dilation: scipy.ndimage.binary_dilation with a large disk structuring element was pathologically slow on one contiguous bright source (16.2s on a synthetic case matching that shape); replaced with a distance-transform threshold, the same result computed near-linearly.
• Post-process hot-pixel median filter: was calling scipy directly instead of this project's own native per-channel median kernel — same class of miss already fixed once elsewhere, missed here. 5.1s -> ~0.1s.
• gaussian_filter_ds's upsample: cubic -> bilinear interpolation (the input just came out of the function's own huge blur, so cubic recovers nothing real). ~3x/call.
• New native Gaussian blur kernel, swept across every direct scipy.ndimage.gaussian_filter call site that uses a plain 2-D scalar sigma (~15 files) — correlate1d was the single largest self-time item in every profile taken. ~4x faster at small sigma, ~1.3-1.5x at large.

Also included

• A separate, unrelated fix from the same session: per-frame cosmic-ray rejection's >=20-frame auto-skip is now sub-length-aware (Config.LACOSMIC_LONG_SUB_EXPTIME_S), not just frame-count-aware — keeps per-frame detection on for long-sub sessions where the noise win is real and the star-softening cost is negligible, instead of always deferring to the frame-count skip.
• Real benchmark refresh: ran tools/bench_vs_siril.py for real against a local Omega Nebula session (Siril CLI installed locally) on this branch's code, found the cosmic-ray fix above now correctly triggers on that session's real 30s subs, and updated README.md/docs/index.html with the real new numbers (113s -> 132s, quieter: noise 0.99-1.19x Siril's -> 0.87-1.03x). Flagged Sunflower/Sculptor as not yet re-measured rather than leaving them silently stale next to freshly-updated Omega figures.
• A real bug found and fixed while sweeping the Gaussian kernel: the wrapper didn't preserve input dtype, silently upcasting float32 callers to float64. Fixed, with a new test pinning it.
• One attempted optimisation (finer-grained rayon parallelism in small_times_wide) was measured, found to make no difference (memory-bandwidth bound, not core bound), and reverted rather than shipped.

Test plan

• Full test suite passes (1647 tests) on the merged result
• ruff check . and tools/lint_conventions.py clean
• Every native kernel has bit-exact or documented-tolerance parity tests vs. its numpy/scipy fallback
• Real end-to-end runs against real Omega Nebula session data (both a synthetic dataset and the user's actual 114-frame session)
• Two of the newly-wired opt-in paths (--trail-reject, --noise-validate) smoke-tested against real data

astro_native bumped 0.32.0 -> 0.34.0. Full narrative with every measurement in CLAUDE.md and CHANGELOG.md.

🤖 Generated with Claude Code