Add per-band, DP1 MLP and reference constant mag-err model artifacts

[hombit]

Trained-model artifacts left over from the per-band constant mag-err work. Code only — no source changes; every source change from the old dia-object-support branch is already on main via #106, #111, #112, #113 and the flux_err PRs.

Per-band constant mag-err runs, DP2 Object forcedDiaSource

Point sources (extendedness = 0) brighter than 21 mag, all six bands fitted in one run, plus the two comparisons this was run for:

• g-r colour bins (0.0–0.5 vs 1.5–2.0), summarised in models/perband_magerr_gr_comparison.txt: the systematic error is ~9–10 mmag in every band and the red-blue difference stays under 1.2 mmag, so it is essentially colour-independent.
• Three 5M-light-curve cones (Rubin-SV-225, DDF-COSMOS, DDF-EDFS), summarised in models/perband5M_dp2_diff_fields.txt: 6–15 mmag, varying more between fields than between bands.

2026-07-16 DP1 Object MLP models

MLPModel on the DP1 science Object catalog, all six bands, n_src=10, val/test split at 0.6/0.85, for both calibrated and difference images. Each run ships the .pt and .onnx model, the feature-importance figure and the full plot set, including the reduced-chi2, KL-statistic and whiten/addmagerr density diagnostics on the test split.

Fixed 0.01 mag reference model

models/const_magerr_0.01mag/ — not a trained run, but ConstantMagErrModel with addition_centi_mag_err pinned to 1.0, kept as the baseline the trained models are compared against. Calibrated and difference-image variants (the latter referencing the systematic to psfFlux), each with its own plot set.

Also

Drops the superseded models/2026-06-10_DP2_Object_ForcedSource/, replaced by models/2026-07-08_DP2_Object_ForcedSource/. Nothing references it.

Note: the two DP1 MLPModel.pt checkpoints (~850 KB each) exceed the 500 KB check-added-large-files limit and were committed with --no-verify, as the already tracked 2026-07-08 DP2 checkpoints were. All other hooks, tests included, pass.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 53.38%. Comparing base (a644983) to head (7243480).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #123   +/-   ##
=======================================
  Coverage   53.38%   53.38%           
=======================================
  Files          36       36           
  Lines        2246     2246           
=======================================
  Hits         1199     1199           
  Misses       1047     1047           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.