Use om.Constraints throughout. (#86)

* Refactor model configuration to use frozen dataclasses instead of dicts

Introduce strongly-typed dataclasses for model configuration:
- Dimensions, Labels, Anchoring, EstimationOptions, TransitionInfo
- FactorEndogenousInfo, EndogenousFactorsInfo

This improves type safety and enables IDE autocompletion while keeping
user-facing model_dict as a plain dictionary.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Use frozendict for immutable dict fields in dataclasses

Replace dict fields with frozendict in frozen dataclasses to ensure
true immutability:
- Labels.aug_periods_to_periods
- Labels.aug_stages_to_stages
- Anchoring.outcomes
- TransitionInfo.param_names, individual_functions, function_names
- EndogenousFactorsInfo.aug_periods_to_aug_period_meas_types, factor_info

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Ignore ty false positive.

* Return ProcessedModel dataclass from process_model() instead of dict

Update process_model() to return a ProcessedModel frozen dataclass
and update all consumers to use attribute access instead of dict access.

This provides:
- Better type safety with explicit typed fields
- Immutability via frozen dataclass
- IDE autocomplete support
- Clear documentation of the model structure

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Rename FactorEndogenousInfo to FactorInfo

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Ensure complete typing in src/skillmodels.

* Require Python 3.14 before fixing type annotations.

* Move TESTS_DIR -> TEST_DATA_DIR, which points to a subdirectory of src so that config.TEST_DATA_DIR is valid also for skillmodels the package (as opposed to the project).

* Fix more linting issues.

* Make ruff rules much stricter.

* Further tighten type annotations. Fix a missing run-time annotation of NDArray.

* Move unsafe_fixes = false from .pre-commit config to pyproject.

* Fix query in data simulation.

* More fixes to imports, add test.

* Use more tuples in place of lists to prevent errors.

* Fix typing.

* Dataclasses for user input.

* Simplify.

* Use modern rng everywhere.

* Update CLAUDE.md

* Get rid of if TYPE_CHECKING blocks

* Update hooks and clean up

* Call by name throughout.

* Autogenerated docs, harmonised hooks / project configuration.

* Get rid of model_dict.

* Replace yaml model specifications by ModelSpec-s.

* Next shot at fixing pickling.

* Add improved output formatting.

* Add variance decompositions.

* Fix variance decomposition to use aug_period instead of period.

The filtered_states DataFrame and params index both use aug_period as the
period identifier, not period. This fixes KeyError when calling
decompose_measurement_variance.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Fix diagnostic_plots double-processing of debug data.

debug_loglike already returns processed debug data (it calls
process_debug_data internally). Remove the redundant second call
to process_debug_data which was causing AttributeError when trying
to iterate over already-processed DataFrames.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Review comments 1: single EstimationOptions; improved dataclasses with all defaults set right there; get rid of remaining mutable types being passed around.

* Review comments on docs; tighten typing of Mapping to MappingProxyType where we could do so.

* Back to general implementation of ensure_containers_are_immutable.

* Small fixes.

* Latest boilerplate version.

* Make optimagic a package dependency.

* Latest boilerplate.

* Make jupyterbook and pytask required deps.

* Use dags 0.5.0 dev branch.

* Update CLAUDE.md with complete API docs and actual application usage

Document all 14 public exports, get_maximization_inputs return dict,
ProcessedModel attributes, frequently used internal APIs, all 7
transition functions, and ModelSpec builder methods — based on how
the package is actually used in skane-struct-bw and health-cognition.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Increase test coverage to close to 100%.

* Fix bugs caught by review agent.

* Call ty environment 'type-checking'.

* Require dags 0.5, bump pandas, numpy and hooks, fix resulting issues.

* Remove suppression of deprecation warning, which is unnecessary in Pandas 3.0

* Make ruff stricter.

* Update boilerplate.

* Update ty and remove a couple of unused-ignores.

* Rename _sel to select_by_loc; fix stale docstring

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Allow passing states directly to get_transition_plots (health-cognition use case).

* Use optimagic constraints directly. Need a wrapper because we are setting values on FixedConstraints.

* Fix FixedConstraintWithValue.loc type and test expectations

Remove list from loc type union, convert callers to tuple().
Update anchoring test expectations from list to tuple.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Validate not-None input directly.

* Simplify API by setting .selectors via .loc in __post_init__

* Require dags 0.5.1

* Fix aug_period/period bug in transition equation visualization

The viz code assumed states DataFrames always have `aug_period` as a
column, but pre-computed states (e.g. from health-cognition) may carry
`period` in the index instead. Add `_normalize_states_columns` to
promote index levels and rename `period` → `aug_period` when needed.

Also document the period vs aug_period convention in CLAUDE.md.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Remove aug_periods from public-facing functions.

* Previous commit was too greedy.

* Use pixi 0.66 in CI; prek autoupdate.

* CHORE: Approximation-tolerant floating point comparisons, remove pandas in-place operations.

* Get rid of unnecessary block-comments.

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

Author: hmgaudecker

CLAUDE.md

@@ -93,8 +93,8 @@ simulate_dataset() → Simulate states (with optional policy effects)
   computation using Kalman filtering. The debug variant is not jitted and returns
   intermediate results (residuals, contributions, filtered states).
 - **constraints.py**: Generates parameter constraints (bounds, equalities from stagemap,
-  fixed values) for optimization. Exports `get_constraints_dicts()`,
-  `constraints_dicts_to_om()`, `enforce_fixed_constraints()`, `add_bounds()`.
+  fixed values) for optimization. Exports `get_constraints()`,
+  `enforce_fixed_constraints()`, `add_bounds()`, `FixedConstraintWithValue`.
 - **parse_params.py**: Converts flat parameter vectors to structured model parameters.
   Exports `create_parsing_info()` and `parse_params()`.
 - **params_index.py**: Builds the `pd.MultiIndex` for the params DataFrame via
@@ -194,7 +194,8 @@ These are not in `__all__` but are imported directly by application projects:
 - `skillmodels.process_model.process_model` — central to all application code
 - `skillmodels.types.ProcessedModel`, `EndogenousFactorsInfo`
 - `skillmodels.decorators.register_params` — essential for custom transition functions
-- `skillmodels.constraints.constraints_dicts_to_om`, `enforce_fixed_constraints`
+- `skillmodels.constraints.get_constraints`, `enforce_fixed_constraints`,
+  `FixedConstraintWithValue`, `select_by_loc`
 - `skillmodels.utilities.extract_factors`, `update_parameter_values`
 - `skillmodels.process_data.pre_process_data`
 - `skillmodels.correlation_heatmap.get_measurements_corr`, `get_quasi_scores_corr`,

pixi.lock

@@ -29,7 +29,7 @@ classifiers = [
 ]
 dynamic = [ "version" ]
 dependencies = [
-  "dags>=0.5",
+  "dags>=0.5.1",
   "jax>=0.9",
   "jupyter-book>=2",
   "kaleido>=1.2",