trace: add max_error to reject fused double-order clusters

[nikolai-piskunov]

Summary

Adds an optional max_error parameter to trace(). After the final polynomial fit, any cluster whose residual RMS (in pixels) exceeds max_error is discarded.

The motivation is closely-packed detectors where two adjacent spectral orders can be merged into a single cluster (typically at detector edges, where the gap between orders closes up). Such a fused cluster cannot be followed by a single polynomial, so its fit RMS is large — on the order of half the order separation — whereas a genuine single-order trace fits to sub-pixel RMS. Thresholding on the fit RMS cleanly rejects the fused clusters while leaving real orders untouched.

This ports the MAX_ERR keyword from the IDL REDUCE mark_orders, where it was added for exactly this case (e.g. MOSAIC NIR, ~5 px median order spacing).

It is disabled by default (max_error=None), so existing reductions are unaffected.

Changes

• trace.trace(): new max_error=None parameter; reject clusters with fit RMS above it (logged when any are dropped).
• reduce.py Trace step: read max_error from config (defaults to None) and pass it through.
• instruments/defaults/schema.json and defaults/settings.json: add max_error (default null).
• test/test_trace.py: TestMaxError — disabled / generous / tiny thresholds, plus a fused high-RMS cluster being dropped while a clean order survives.

Notes

While porting I checked the other order-tracing improvements from the IDL side and found PyReduce already covers them: the 90 % overlap auto-merge guard (auto_merge_threshold), degree-2 fitting before merge (degree_before_merge), and gap handling in determine_overlap_rating. numpy's SVD-based lstsq is also robust to the polynomial conditioning that needed an explicit basis rescale in IDL — verified no RankWarning and identical coefficients at degree 4 over 4096 columns. So max_error is the one genuinely new piece.

Testing

pytest test/test_trace.py and test/test_configuration.py pass (66 trace tests including the 4 new ones; config tests confirm the default settings still validate against the schema).

🤖 Generated with Claude Code

[ivh]

Merged into master via rebase (so GitHub doesn't auto-mark this as merged). Landed as commits cf81ab4 (the max_error feature) and 977a0cb (added fused-order + config plumbing tests). Both are on origin/master.