fix: break the calibration chain's dependency on its last step

.claude/skills/model-calibration/SKILL.md

@@ -299,6 +299,14 @@ historical centre once the chain absorbs its gaps cleanly.
   a lower triangle (feed disables the other four, food_waste the last
   three, and so on); preserve it when adding a calibration.
 
+  The trap is that a step need not *enable* a successor's calibration to
+  depend on it -- inheriting `l1_cost: "calibrated"` from `default.yaml` is
+  enough. feed/food_waste/food_demand therefore set
+  `deviation_penalty.enabled: false` (production is pinned to actuals, so the
+  penalty has nothing to act on) and cost drives stability through hard
+  bounds. When adding a calibrated key to `default.yaml`, check what the
+  earlier step configs inherit.
+
 - **Step configs must not override structural model keys.** Anything a
   step config pins that isn't calibration machinery, the validation
   regime, or a solver budget silently defeats the point of a per-base

CHANGELOG.md

@@ -261,6 +261,16 @@ introduce breaking changes to configuration and outputs.
 
 ### Fixed
 
+- The feed, food_waste, food_demand and cost calibration steps no longer depend
+  on the calibrated deviation penalty, which the stability step produces at the
+  end of the chain. The first three pin production to actuals, so the penalty
+  had nothing to act on; the cost step drives production stability through hard
+  bounds, which never read the calibrated L1 costs. Regenerating the artefact
+  sets confirms the dependency was inert: every artefact is unchanged. The
+  calibration chain is a strict forward pass again, and
+  `tools/calibrate --check` settles after a full run instead of reporting the
+  first four steps stale.
+
 - Workflow startup no longer builds a full copy of the configuration for every
   configured scenario when deciding whether health data is needed. On configs
   with generated scenario ensembles this dominated DAG construction: for

config/calibration/feed.yaml

@@ -69,6 +69,12 @@ food_demand_calibration:
 food_loss_waste_calibration:
   enabled: false
 
+# Production is pinned to actuals here, so the deviation penalty has nothing
+# to act on. Leaving it enabled would make this step read the calibrated L1
+# costs, which the stability step produces at the end of the chain.
+deviation_penalty:
+  enabled: false
+
 grazing:
   grassland_forage_calibration:
     enabled: false     # Don't apply calibration (we're generating it)

config/calibration/food_demand.yaml

@@ -66,6 +66,12 @@ food_utility_piecewise:
 cost_calibration:
   enabled: false
 
+# Production is pinned to actuals here, so the deviation penalty has nothing
+# to act on. Leaving it enabled would make this step read the calibrated L1
+# costs, which the stability step produces at the end of the chain.
+deviation_penalty:
+  enabled: false
+
 food_demand_calibration:
   enabled: false     # Don't apply calibration (we're generating it)
   generate: true     # Generate calibration from solved model

config/calibration/food_waste.yaml

@@ -57,6 +57,12 @@ food_utility_piecewise:
 cost_calibration:
   enabled: false
 
+# Production is pinned to actuals here, so the deviation penalty has nothing
+# to act on. Leaving it enabled would make this step read the calibrated L1
+# costs, which the stability step produces at the end of the chain.
+deviation_penalty:
+  enabled: false
+
 # food_demand calibration depends on food_loss_waste.csv, which is built
 # from this step's output. Disable to break the cycle: food_waste's slack
 # must reflect the gap *without* the demand multiplier applied (the

data/curated/calibration/default/deviation_penalty.yaml

@@ -17,6 +17,6 @@ l1_costs:
   cropland: 1.4406072051746193
   grassland: 0.23976429357967677
   feed: 0.06151640376563935
-iterations: 1
+iterations: 0
 converged: true
 final_residual_log_inf: 0.019950027731896344

data/curated/calibration/default/provenance.yaml

@@ -4,8 +4,8 @@
 # tools/calibrate; do not edit by hand.
 # Licensing: see the annotation in REUSE.toml.
 base_config: config/default.yaml
-generated_at: '2026-07-28T17:44:34+00:00'
-git_commit: e62c3a149e0bc4a9b103ae49c6af12b9e62a254f
+generated_at: '2026-07-28T22:22:47+00:00'
+git_commit: 53225fd85a4fa4ec491f92515226a9d09f790073
 source: default
 structural_config:
   aggregation.irrigated_area_source: current

data/curated/calibration/gbd-anchored/deviation_penalty.yaml

@@ -17,6 +17,6 @@ l1_costs:
   cropland: 1.606710895127586
   grassland: 0.24247737358625074
   feed: 0.05417951839730158
-iterations: 1
+iterations: 0
 converged: true
 final_residual_log_inf: 0.018045776200625075

data/curated/calibration/gbd-anchored/provenance.yaml

@@ -4,8 +4,8 @@
 # tools/calibrate; do not edit by hand.
 # Licensing: see the annotation in REUSE.toml.
 base_config: config/gsa.yaml
-generated_at: '2026-07-28T18:00:37+00:00'
-git_commit: e62c3a149e0bc4a9b103ae49c6af12b9e62a254f
+generated_at: '2026-07-28T22:30:03+00:00'
+git_commit: 53225fd85a4fa4ec491f92515226a9d09f790073
 source: gbd-anchored
 structural_config:
   aggregation.irrigated_area_source: current

workflow/rules/common.smk

@@ -327,9 +327,16 @@ def calibration_artefact_inputs(cfg):
     if fd_cal_cfg["enabled"]:
         inputs["food_demand_calibration"] = fd_cal_cfg["calibration_file"]
 
+    # Mirrors the solve's own gate: the sentinel is resolved only for an
+    # enabled L1 penalty, so anything else never opens the file.
     dp_cfg = cfg["deviation_penalty"]
     dp_cal_cfg = dp_cfg["calibration"]
-    if dp_cal_cfg["enabled"] and deviation_penalty_uses_calibrated(dp_cfg):
+    if (
+        dp_cal_cfg["enabled"]
+        and dp_cfg["enabled"]
+        and dp_cfg["penalty_mode"] == "l1"
+        and deviation_penalty_uses_calibrated(dp_cfg)
+    ):
         inputs["deviation_penalty_calibration"] = dp_cal_cfg["calibrated_yaml"]
 
     return inputs