[lachlangrose]

Fixes from a code review of the stack (#132 to #148). This PR is on top of #148.

Changes

• Derived refresh: take the layer role first, then the tool setting. This matches the inputs that the refresh records, so a result is not marked current for a layer it did not use.
• Native solver: forget the fallback state of a feature when its data is read again or the model is rebuilt. _needs_inequalities is built again on a data refresh.
• Added constraints: skip a row whose layer was removed from the project, and log it, instead of failing the update.
• Parametric faults: spec_from_dict raises ValueError for a damaged spec. The project load skips that fault and logs it. Removing a fault clears its state and emits model_updated.
• Model step: changed layers are a to-do, not a problem. The text names the buttons that exist.
• Export: a NaN id has no unit (-1). The CSV is written with the csv module, so a quote in a unit name is escaped.

Not changed

• summarise_data counts a row with both val and gradient columns twice. Both are real constraints, so this is correct.
• The release-please manifest changes came from release-please commits, not a hand edit.

Tests

• 351 unit tests pass. Black and ruff pass on the changed files.
• Two tests checked the old behaviour and are updated. New tests cover the CSV quote, the NaN id and damaged parametric specs.
• The QGIS tests (tests/qgis) were not run.