N_EPOCH counts only PSF-validated epochs

[cailmdaley]

The final catalogue's N_EPOCH counts epochs with a validated PSF (point-spread function), the epochs available to ngmix.
There is one coverage-count column: the geometric overlap count remains internal and is not written to the final catalogue.
This PR is stacked on #942, with base feat/final-cat-hdf5.

PSF validation and epoch counts

tile_detect writes the sexcat's geometric N_EPOCH before any PSF exists, by projecting each object into exposure CCD footprints.
psfex_interp only interpolates on CCDs passing validation (ACCEPTED ≥ STAR_THRESH = 22, CHI2 ≤ 2), and ngmix uses those PSF vignets.
Previously, the final catalogue retained the geometric count even when a covering CCD failed validation.

On nibi's smk-g16 (25 tiles, 813,009 objects), 0.6–1.3% of objects have a +1 discrepancy against the PSF-validated exposure-count maps from #887.
Most form whole CCD footprints.
On tile 188.308, exposure 2603246 CCD 35 has CHI2 0.79 but only 21 accepted stars:

This change removes the +1 discrepancy caused by PSF rejection.
The −1s at CCD edges come from geometry: the map includes DATASEC's half-pixel border, while the catalogue uses strict bounds.
That convention is unchanged.

Catalogue design

• make_cat overwrites N_EPOCH with the number of entries in each object's galaxy-PSF catalogue, or 0 for "empty".
  An epoch with a failed HSM PSF-shape fit still counts because its PSF model exists.
• N_EPOCH is written even without SAVE_PSF_DATA; per-epoch slot columns still follow that setting.
• Default slot sizing uses the sexcat's geometric max(N_EPOCH) + 1 as an upper bound, before overwriting the count.
  Explicit N_EPOCH_SLOTS is unchanged.
• The final catalogue and parameter files contain no N_EPOCH_OVERLAP.
  PSF-gate rejections are studied per CCD through validation output and Exposure-level HealSparse maps: exposure count and flagged pixels #887 exposure maps, rather than through a second per-object count.
  One column avoids applying downstream cuts to the wrong count.
• astra.yaml records this choice in epoch_membership_ccd_bounds and psf_acceptance_thresholds; their committed options are unchanged.

sp_validation's N_EPOCH >= 2 cut now requires two PSF-validated epochs.
An object whose second covering CCD has no validated PSF no longer passes that cut.
ngmix can reject additional epochs, so NGMIX_N_EPOCH <= N_EPOCH.

Catalogues made before this change

For catalogues written with SAVE_PSF_DATA, migrate in place by setting N_EPOCH := count of EXP_ID_n != -1 across the per-epoch slots.
No additional column is created, and no pipeline rerun is needed.
Catalogues without those slots need the galaxy-PSF catalogue to reconstruct the validated count.

Checked on nibi production wave 1 (b01–b05, 125 tiles, 5,015,937 objects): NGMIX_N_EPOCH <= N_EPOCH <= old N_EPOCH for every object; 99.42% unchanged.

Validation

In the develop SIF (shapepipe-develop-240b37e4.sif), with this checkout's src/ shadowed through PYTHONPATH:

• 180 tests passed across make_cat, make_cat_mask_ext, hsm_column_seams, psf_grammar_properties, final_cat_merge_invariants, final_cat_reader, mask_ext_columns, decisions, and ngmix_range.
• The runner regression checks sexcat counts [2,3,1] become final N_EPOCH [1,3,0], with no overlap column, even without per-epoch slots.
• uvx astra-tools@0.2.17 validate passed for astra.yaml and universes/committed.yaml.
• The workflow parameter-pin test could not run in this SIF because snakemake is not installed.

GPT-6.1 Sol via Claude Opus on behalf of Cail.