Adopt upstream radiometric calibration QC filter and retrain

[thomasteisberg]

Summary

The upstream radar-return-statistics stores now ship per-trace radiometric calibration diagnostics (method 0.4.1: image-combine seam offsets, surface source image, surface ceiling margin). This PR adopts the changelog's suggested QC filter, re-pins all three stores to the calibration snapshots, retrains both required-SNR models, and documents the effect.

• split.calibration_qc (new, default on): reject traces with a seam step ≥ 3 dB, a surface sampled from a higher-gain image, or a surface within 2 dB of its season's clip level. Rules pass where unmeasured, because the seam check could not run on 92% of 2014_Greenland_P3 and 61% of 2012_Antarctica_DC8. Drops 14% of Antarctic and 13% of Greenland traces after the existing season exclusions.
• Re-pinned snapshots to the 2026-09-03 method-0.4.1 refresh (antarctica 087JBD7NTAE8BTBTEYSG, greenland ACA8WY61ZF9W6VSBA3HG, ase F6TXWSEQ9RQCD1H7MSMG); upstream verified no pre-existing values changed, so old vs new differs only by the filter. The 0.4.1 refresh over 0.4.0 changed nothing that reaches the model (72 already-rejected 2018 DC8 traces lost an at-ceiling flag); the retrain reproduces the 0.4.0 posteriors exactly.
• Retrained models (run_ids atten_refl af5514ae3b72, linear c0b385764662): atten_refl CV RMSE 13.02 → 12.87 dB, test 12.72 dB; linear 13.92 → 13.80 dB. Mission tool payload regenerated.
• Snakefile: each rule depends on a stamp of exactly the config section its run_id hashes (outputs/config_stamps/), so a re-pin or config edit re-runs only the stages it affects.
• Comparison tooling: scripts/qc_filter_comparison.py, season_crossover_matrix.py --calibration-qc.

Assessment

The filter is sound hygiene with almost no effect on the model. Scored on the same held-out points, pre- and post-QC posteriors differ by 0.03 dB RMSE; every parameter stays within ~2 posterior sd; full-grid predictions shift by −0.2 to +0.6 dB (5th–95th pct). Model-free crossovers show the one real repair: 2016_Greenland_P3 (49% seam rejects) now agrees with the other P3 seasons (offsets ~0 dB, scatter 12 → 8 dB). Season-level offsets (2012 DC8 ~12 dB low, 2017 Basler ~30 dB low, 2013 Greenland 16–19 dB low) are untouched, so exclude_collections stays. Full write-up: agent_notes/20260901-calibration-qc-results.md; docs updated in docs/2_input_data.md and docs/3_model.md.

Test plan

• uv run pytest -q (68 passed, incl. new tests/unit/test_calibration_qc.py)
• uv run ruff check src tests scripts
• uv run snakemake --cores 4 model_all end to end on the re-pinned stores (0 divergences, R̂ ≤ 1.005), twice: 0.4.0 and the 0.4.1 refresh, bit-identical posteriors
• Section-specific rerun triggers verified with scripted config edits (see commit ec48bfb)

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y73QhGDAtbDwo9ZnJ2N4F2

[thomasteisberg]

Good catch, fixed in ec48bfb (see the commit for details). Rather than adding model.yaml as a grid input (which would also rebuild the slow grid on any train/split edit), each rule now depends on a stamp of exactly the config section its run_id hashes, written by the Snakefile at parse time only when that section's hash changes:

• grid: inputs + grid
• split: inputs + split
• train: train + that model's entry (one stamp per model)
• augment: the store's full config (replaces the whole-file input, so comment-only edits and re-pin annotations no longer rerun a 10-minute extraction)

First-time stamps get an epoch mtime, so migrating existing outputs does not trigger a rebuild. Verified by dry runs after scripted edits: grid edit → grid + split + train; split edit → split + train (no grid); linear-only entry edit → linear only; draws edit → both models; comment-only edits in either config → nothing; Greenland min_thickness_m edit → Greenland augment onward.

[thomasteisberg]

Re-pinned to the 0.4.1 refresh (2026-09-03 15:51 UTC snapshots) and retrained in 6a1fb54. The refresh's only model-visible change is 72 traces of 2018_Antarctica_DC8 losing their at-ceiling flag (its rising img2 envelope no longer counts as a ceiling), and all 72 were already rejected by the img2 rule, so the split is identical apart from run_ids and both models retrain to bit-identical posteriors. Only provenance (snapshot IDs, run_ids, mission-tool meta) and the doc references changed; the figures did not need regenerating. The config stamps triggered both augment reruns on their own.