image-sims workflow: campaign hardening + snakemake-native containers

src/sp_validation/image_sims.py

@@ -152,11 +152,17 @@ def load_catalogs(self, verbose=True):
         grids_dir = self.cfg["grids_dir"]
         num = self.cfg["num"]
         cat_name = self.cfg.get("catalog_name", "shape_catalog_cut_ngmix.fits")
+        # Sim-directory naming mirrors the workflow's contract: grid sims live in
+        # ``{branch}_grid_{num}``, all other families (e.g. blended) in
+        # ``{branch}_{num}``.  Defaults to ``grid`` so pre-``sims_type`` configs
+        # and the synthetic-recovery tests keep the historical layout.
+        sims_type = self.cfg.get("sims_type", "grid")
+        suffix = f"_grid_{num}" if sims_type == "grid" else f"_{num}"
         # ``sim_names`` (incl. the unsheared reference) comes from the config's
         # branch map. The +g/-g pool estimator pairs the sheared sims directly;
         # the reference is loaded for completeness and null-test diagnostics.
         for name in self.sim_names:
-            path = f"{grids_dir}/{name}_grid_{num}/{cat_name}"
+            path = f"{grids_dir}/{name}{suffix}/{cat_name}"
             if verbose:
                 print(f"  Loading {path}")
             self.cats[name] = _load_cat(path, self.w_cols)

workflow/image_sims/Snakefile

@@ -33,9 +33,10 @@ there under ``if "image_sims" in config``.
 configfile: "workflow/image_sims/config.yaml"
 
 
-# The image-sims rules own their container invocation explicitly, so no
-# top-level container is needed here.
-container: None
+# Each rule declares its own image via ``container:`` (the ShapePipe vs
+# sp_validation split); with ``--software-deployment-method apptainer`` in the
+# candide profile, Snakemake wraps every job in ``apptainer exec``.  No
+# top-level container here -- the per-rule directives own the choice.
 
 
 include: "../rules/image_sims.smk"

workflow/image_sims/config.yaml

@@ -51,6 +51,13 @@ image_sims:
   # --- simulation grid --------------------------------------------------
   sims_type: grid        # 'grid' -> *_grid_{num}; anything else -> *_{num}  [operational]
   num: 1
+  # Optional: pull exposures from a *different* realization than tiles. Unset ->
+  # exposures follow `num` (prior behaviour, bit-for-bit). Only meaningful for
+  # non-grid sims: tiles stay on `{branch}_{num}`, exposures come from
+  # `{branch}_{exp_num}`. Reproduces Fabian's blended production split -- the
+  # blended-family realizations _1.._4 carry stale cross-linked SP_exp (grid-sims
+  # exposures symlinked in, signal-free); _5 is the clean all-real exposure set.
+  # exp_num: 5
   # Branches this run requests: the unsheared reference plus the four +/-
   # sheared branches.  Injected shear is NOT set here -- it is parsed from each
   # branch's basic_info.txt by im_manifest.  [operational]
@@ -63,6 +70,13 @@ 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

@@ -11,29 +11,30 @@
 # cluster policy (executor, account, partition, node excludes, per-job
 # defaults); it carries no science and no workflow logic.
 #
-# What is deliberately NOT here:
-#
-#   * Container / apptainer settings.  The image-sims rules set
-#     ``container: None`` and own their ``apptainer exec`` call through the
-#     shared ``EXEC`` prefix (one image for every stage, with PYTHONPATH /
-#     PSF_DICT / OMP_NUM_THREADS injected there).  So no
-#     ``software-deployment-method: apptainer`` / ``apptainer-args`` -- those
-#     would wrap a *second*, redundant container around jobs that already run
-#     inside one.
+# Containers: the image-sims rules declare their image with Snakemake's own
+# per-rule ``container:`` directive (the ShapePipe vs sp_validation split), and
+# ``software-deployment-method: apptainer`` here turns that into an
+# ``apptainer exec <image>`` around every job.  ``apptainer-args`` carries the
+# static bind mounts -- the candide paths every job needs mounted.  The per-run
+# env injections (PYTHONPATH shadowing the branch build, PSF_DICT, OMP_NUM_THREADS
+# =1) are NOT here: they are per-run values derived from the run config's repo and
+# PSF paths, so they live in committed workflow code (``os.environ`` in
+# workflow/rules/image_sims.smk) and reach the container natively -- the slurm
+# executor submits ``sbatch --export=ALL`` and apptainer inherits the host env by
+# default.  Setting them in the .smk, not by hand in the launching shell, keeps
+# the "one run command, no implicit uncommitted state" property.
 #
-#   * OMP_NUM_THREADS.  It is pinned to 1 on the ``apptainer exec`` line in
-#     workflow/rules/image_sims.smk, not here.  The slurm executor submits with
-#     ``--export=ALL``, which propagates the *driver's* ambient environment; a
-#     profile only sets CLI flags, never the driver's own env, so an
-#     ``OMP_NUM_THREADS`` set here would silently depend on the operator having
-#     exported it by hand.  Injecting it at the container boundary puts it where
-#     the compute runs, committed and independent of the launching shell.
+# What is deliberately NOT here:
 #
 #   * Per-rule resources (mem_mb, runtime).  Those live on each rule in the
 #     .smk; the ``default-resources`` below are only the floor for rules that
 #     set none.
 
 executor: slurm
+software-deployment-method: apptainer
+# Bind the candide filesystems every job's inputs/outputs/repos live on.
+# /automnt is required when repos or data resolve through the automounter.
+apptainer-args: "--bind /n17data,/n09data,/home,/automnt"
 
 # 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,