Calibration fail-fast, pipeline_config_dir, jobs cap, response docs (imsims A/B campaign)

src/sp_validation/calibration.py

@@ -40,6 +40,16 @@ def get_calibrated_quantities(gal_metacal):
         weights
     mask : array of bool
         mask to indicate valid objects in "no-shear" sample
+
+    Notes
+    -----
+    Calibration uses a single global response: ``gal_metacal.R`` is the
+    2x2 shear+selection response matrix averaged over the whole run, and
+    the same matrix is applied to every object. Per-object responses
+    (``R_g11`` ... ``R_g22``) are computed upstream but are only ever
+    consumed as catalogue-wide (or, in ``get_w_des`` /
+    ``get_quantities_binned``, per-bin) means -- no object is calibrated
+    with its own response.
     """
     # mask for 'no shear' images
     mask = gal_metacal.mask_dict["ns"]
@@ -976,6 +986,14 @@ def _masking_gal(self):
 
             self.mask_dict[name] = ind_masked
 
+        if len(self.mask_dict["ns"]) == 0:
+            raise ValueError(
+                "0 objects survive the metacal selection (no-shear sample"
+                " is empty: flag == 0 + SNR + size cuts left nothing) --"
+                " upstream shape measurement likely failed wholesale;"
+                " check the ngmix outputs before calibrating."
+            )
+
     def _masking_gal_mom(self):
         """Add docstring.
 

workflow/image_sims/Snakefile

@@ -17,8 +17,10 @@ command on candide, with the committed SLURM profile owning all scheduling:
 defaults automatically, so only ``my_run.yaml`` (the science knobs, and any
 operational override that run wants) is passed on the command line; Snakemake
 deep-merges the two.  The profile supplies the executor, account, partition,
-node excludes and job floor -- no ``-j`` needed (the slurm executor sets the
-job cap).  Off-cluster, drop ``--profile`` and add ``-j N`` to run locally.
+node excludes, job cap and job floor -- no ``-j`` needed on the command line
+(snakemake >= 9 requires an explicit cap even with the slurm executor, so the
+profile carries ``jobs``).  Off-cluster, drop ``--profile`` and add ``-j N``
+to run locally.
 
 The target (``im_mbias``) is given *before* ``--configfile``: Snakemake's
 ``--configfile`` takes one-or-more paths, so a target placed after it is

workflow/image_sims/config.yaml

@@ -63,6 +63,13 @@ image_sims:
   # ShapePipe cfis configs (final_cat.param etc.); default is
   # {shapepipe_repo}/example/cfis_image_sims once #766 lands.  [operational]
   config_dir: /n17data/cdaley/unions/scratch-wf/imsims-run/grids/_cfis_image_sims
+  # ShapePipe config tree the *pipeline* stage runs against (run_job's -c
+  # override), the one that owns the ngmix ini and hence METACAL_PSF.  Empty ->
+  # run_job self-computes {shapepipe_repo}/example/cfis_image_sims (the
+  # historical default, byte-exact).  Point it at a per-arm copy carrying a
+  # different ngmix ini to A/B the metacal PSF without editing the worktree.
+  # [operational]
+  pipeline_config_dir: ""
   psf_model: psfex       # [operational]
   n_smp: -1              # [operational]
   # Extract/calibrate scripts run from the sp_validation repo checkout (branch

workflow/profiles/candide/config.yaml

@@ -35,6 +35,11 @@
 
 executor: slurm
 
+# Snakemake >= 9 requires an explicit job cap even with the slurm executor
+# (9.16.3 errors out without one); carrying it here keeps the run command
+# free of ``-j`` as the Snakefile docstring promises.
+jobs: 100
+
 # Cluster policy applied to every job unless a rule overrides it.  The excludes
 # are the flaky/no-internet candide nodes (n17 mount issues, n09 no internet,
 # n36); ``slurm_extra`` is passed verbatim onto the sbatch line by the executor

workflow/rules/image_sims.smk

@@ -86,6 +86,7 @@ _OPERATIONAL_KEYS = {
     "branches",
     "shape",
     "config_dir",
+    "pipeline_config_dir",
     "psf_model",
     "n_smp",
     "extract_script",
@@ -168,6 +169,13 @@ MASK_CONFIG = IMSIM["mask_config"]  # e.g. config/calibration/mask_v1.X.9_im_sim
 PARAMS_TEMPLATE = f"{SPV_REPO}/workflow/image_sims/params_im_sim.py"
 # ShapePipe cfis_image_sims config dir (per-tile/exposure configs + final_cat.param).
 CONFIG_DIR = IMSIM["config_dir"]
+# ShapePipe config dir the *pipeline* stage's run_job uses (its ``-c`` override,
+# the tree that owns the ngmix ini and thus METACAL_PSF).  Empty (the default)
+# means "let run_job self-compute {shapepipe_repo}/example/cfis_image_sims" --
+# byte-for-byte the historical behaviour, so an unset key reproduces prior runs
+# exactly.  Set it (e.g. per A/B arm) to point the ShapePipe stage at a config
+# tree carrying a different ngmix ini without touching the shared worktree.
+PIPELINE_CONFIG_DIR = IMSIM["pipeline_config_dir"]
 
 # ShapePipe scripts live in the ShapePipe repo (also baked into its image).
 CREATE_FINAL_CAT = f"{SHAPEPIPE_REPO}/scripts/python/create_final_cat.py"
@@ -372,14 +380,18 @@ rule im_pipeline:
         run_dir=lambda wc: f"{GRIDS_BASE}/{wc.sim}",
         psf=IMSIM["psf_model"],
         n_smp=IMSIM["n_smp"],
+        # ``-c DIR`` only when pipeline_config_dir is set; empty -> run_job
+        # self-computes its default, so the flag (and its output) is absent for
+        # a bit-exact reproduction of a run that predates this key.
+        config_flag=f"-c {PIPELINE_CONFIG_DIR}" if PIPELINE_CONFIG_DIR else "",
     resources:
         mem_mb=16000,
         runtime=720,
     shell:
         "cd {params.run_dir} && "
         "{EXEC_PIPELINE} bash {RUN_JOB} "
         "-e {wildcards.tile} -t image_sims -j {JOB_MASK} "
-        "-p {params.psf} -N {params.n_smp}"
+        "-p {params.psf} -N {params.n_smp} {params.config_flag}"
 
 
 rule im_merge: