S1: Read regression_targets in the catalog; migrate the 46 fixture sites #86

Epic: #85 · Order: Story 1 of 8 · Gate: blocks S2, S3, S6 · Closes: #84

Background

views_baseline/model/catalog.py reads self.config["targets"] at 7 sites — lines 57, 62, 67,
75, 85, 98, 112, one per _get_* builder. views-pipeline-core retired that key on 2026-08-02
(507ae11, refs #380/#381, shipped in 3.0.0); the replacement vocabulary is regression_targets
(+ classification_targets, which no baseline uses). Every baseline model has raised
KeyError: 'targets' since. See #85 for the full incident.

The reason the suite did not catch it is tests/conftest.py:12-13, which supplies both keys:

"targets": ["synth_target"],
"regression_targets": ["synth_target"],

That fixture is not merely stale — it is illegal upstream. pipeline-core's combined_targets()
raises ValueError on any config carrying targets, and the manager's config reaches the forecast
stage (manager.config is a property aliasing self.configs, passed as context.configs and read
by combined_targets(context.configs) at forecasting/stage.py:153,173). A fixture that pipeline-core
would reject is not a realistic merged config, and leaving it in reproduces the exact blind spot
that caused this outage.

Work

1. views_baseline/model/catalog.py — derive the target list once in __init__:

   def __init__(self, config: dict, partition_dict: dict, loa: str):
       """Catalog of available baseline models.
   
       Reads the views-pipeline-core config vocabulary: `regression_targets`.
       (`targets` was retired in pipeline-core #380 and is now rejected by
       `combined_targets()`.) Classification targets are not read — no baseline
       is wired for them; see ADR-012 when that changes.
       """
       self.config = config
       self.partition_dict = partition_dict
       self.loa = loa
       self.targets = list(config["regression_targets"])
       ...

   Then replace all seven targets=self.config["targets"] with targets=self.targets.

   Do not import combined_targets — model/ may not import views-pipeline-core at module
   level (ADR-013). See [EPIC] Restore baseline runnability — migrate off the retired targets config key #85 for why the manager-passes-it-down variant was also rejected.

2. tests/conftest.py — delete line 12 ("targets": ["synth_target"],). Line 13
   ("regression_targets") already exists and stays. MANAGER_BASE_CONFIG is then the only
   already-compliant fixture in the suite.

3. Migrate the 46 config-key sites. Rename the dict key "targets" → "regression_targets" in:

   • tests/test_catalog.py — 17 configs (lines 14, 31, 45, 59, 74, 91, 101, 111, 123, 144, 162,
     179, 191, 203, 215, 241, 276)
   • tests/test_baseline_manager.py — 11 configs (29, 55, 87, 120, 272, 285, 310, 336, 374, 403, 432)
   • tests/test_reproducibility_gate.py — 12 inert configs (37, 48, 64, 79, 97, 115, 126, 139, 205,
     219, 231, 244) plus the manager config at 172
   • tests/test_baseline.py — 5 configs (572, 581, 591, 600, 609)
   • tests/test_falsification_seed_wiring.py — 2 configs (38, 76)

4. Do NOT touch the ~115 non-config uses. These are a different thing that happens to share the
   word:

   • the targets pytest fixture — tests/conftest.py:73-75
   • model constructor kwargs — ZeroModel(targets=...) etc., ~50 sites in test_baseline.py
     plus test_golden.py, test_parametric.py, test_pooling.py, test_dual_input.py,
     test_grid.py, test_helpers.py, test_closeness.py, test_ensemble_consumption.py
   • to_feature_frame(..., targets=...) — throughout test_input.py, test_pooling.py
   • tests/test_protocol.py:13 — _BASE is a **kwargs bag for direct model construction
     (it contains partition_dict and loa, which no real config does). Leave it alone.
   • parameter names — _frame_config(targets=…) at test_baseline_manager.py:115,
     make_dummy_ff(targets=…) at conftest.py:59, _pandas_pool_reference(…, targets, …)

5. Five assertions read both sides — tests/test_catalog.py:39, 53, 67, 82, 133 assert
   model.targets == config["targets"]. The model attribute keeps its name; only the config
   subscript changes.

Order within the PR (this is the point of the story)

Do step 2 and 3 first and run the suite. It must go red. That red is the evidence the tests
now describe the config shape production actually has. Only then apply step 1 and watch it go green.
A PR that was never red at this commit has not proved anything.

Acceptance criteria

• grep -rn 'config\["targets"\]' views_baseline/ returns nothing
• grep -rn '"targets"' tests/ returns nothing (all remaining uses are targets= kwargs or the fixture)
• grep -n 'targets' tests/test_protocol.py still shows _BASE unchanged
• conda run -n views_pipeline python -m pytest -q → 248 passed, 1 xfailed
• ruff check . clean
• The PR description records that the suite was red between the fixture commit and the catalog commit

Validation

conda run -n views_pipeline python -m pytest -q
ruff check .
# the fix, proven directly:
conda run -n views_pipeline python -c "
from views_baseline.model.catalog import BaselineModelCatalog
c = BaselineModelCatalog(config={'regression_targets':['lr_os_best'],'window_months':18,'lambda_mix':0.05,'n_samples':256,'seed':42}, partition_dict={'test':(493,540)}, loa='pgm')
m = c.get_model('MixtureBaseline'); print('OK', m.targets)
"

Relevant files

views_baseline/model/catalog.py, tests/conftest.py, tests/test_catalog.py,
tests/test_baseline_manager.py, tests/test_reproducibility_gate.py, tests/test_baseline.py,
tests/test_falsification_seed_wiring.py

Labels: story, implementation, bug