Outputs

A frame run (run)

In out_dir:

File	Contents
report.html	Verdict and reasons, notes, DD sills with intervals, velocity by class, figures, requirements traceability.
metrics.json	Everything in the report: verdict, verdict_reasons, notes, double_difference (raw, calibrated, baseline), sill_change_vs_raw, sill_change_vs_baseline, velocity_los, velocity_los_gate, velocity_by_class, velocity_by_class_raw, golden_epoch_check, station counts, the config hash and package version.
stations.csv	Per station: position; MIDAS rates and σ; pixel row/col; LOS vector; GNSS LOS rate; InSAR raw and calibrated at the pixel (pix::) and window median (med::); class.
station_series_mm.csv	Calibrated LOS series at each station (dates × stations).
calibrated_velocity_mm_yr.tif	Calibrated LOS velocity on the DISP grid.
sill_draws.npz	Bootstrap draws of every sill.
figures/semivariogram.svg	DD semivariogram, raw, calibrated and baseline.
figures/velocity_scatter.svg	InSAR against GNSS LOS velocity per station.
figures/class_rmse.svg	RMSE by station class.

In cache_dir (shared by every run on the frame): gnss_midas_enu.csv, gnss_enu_daily.csv.gz, and per stack stack_<key>/ with the raw velocity, the mean temporal coherence and the raw station series.

A VLM run (vlm)

metrics_vlm.json, pair_residuals.csv, station_rates.csv.

A summary (summary)

One HTML page: a row per run linking to its report.html, and the requirements traceability across runs.