make_cat: objects ngmix never fit get NGMIX sentinels but MCAL_FLAGS = 0

make_cat.py の474、494、501行付近から始め、_save_ngmix_data が ngmix の出力カタログにない行をどのように埋めるかを追跡してください。フィットされていない場合の NGMIX_MCAL_FLAGS のデフォルト値がゼロ以外である一方、実際の ngmix の行ではそれをクリアできることを確認してください。final_cat_smk-g7.hdf5 の受け入れ出力を使い、NGMIX_N_EPOCH > 0 と併せて、結果として生じるフラグの挙動を確認してください。

_save_ngmix_data pre-fills every NGMIX_* column with sentinels and overwrites them only for objects found in the ngmix output catalogue. An object ngmix never fit is absent from that catalogue (it raised "0 epoch to process", ngmix.py:2034, or any other exception caught at ngmix.py:1058-1063, and was skipped with continue), so it keeps every default:

NGMIX_G1/G2 = -10          make_cat.py:494
NGMIX_T, NGMIX_FLUX = 0    make_cat.py:474
NGMIX_MCAL_FLAGS = 0       make_cat.py:501
NGMIX_N_EPOCH = 0          make_cat.py:453

MCAL_FLAGS, the column meaning "the metacal fit failed", therefore reads clean for objects that were never fit at all. In final_cat_smk-g7.hdf5 (the 64-tile acceptance run of #879/#886/#887/#880) this is 18,983 of 1,851,100 objects (1.03%): cutting on MCAL_FLAGS == 0 alone gives mean e1 = −0.107 (std 1.03), against −0.004 (std 0.22) once NGMIX_N_EPOCH > 0 is also required.

Minimal fix: initialise NGMIX_MCAL_FLAGS at make_cat.py:501 to a nonzero "not fitted" value instead of 0, so the flag column is honest by default and only a real ngmix row can clear it. A dedicated bit would be preferable to a bare nonzero if the column gains a documented bit table. sp_validation's classification_galaxy_ngmix happens to reject these rows today through its NGMIX_G1_PSF_ORIG_NOSHEAR != -10 guard — an exact float comparison against a sentinel — and CosmoStat/sp_validation#343 now carries an explicit NGMIX_N_EPOCH > 0 guard instead.

— Claude (Fable) on behalf of Cail