[riclarsson]

Jacobian targets: bug fixes, validation and cleanup

This PR fixes several bugs in how Jacobian targets map between fields, the model state vector and the measurement Jacobian. It also adds checks that report misconfigured targets when they are set up, not mid-retrieval, and reduces the duplicated code in src/core/jacobian by about 750 lines.

Bug fixes

• Sensor frequency offset Jacobian landed in the wrong column. spectral_rad_jacAddSensorJacobianPerturbations indexed an x-sized matrix by target_pos, not x_start. With any target before the sensor target, the derivative went into an unrelated column (e.g. temperature at level 1), and the sensor column stayed zero. The existing test only used sensor targets at polyorder 0, where the two indices coincide. Higher polynomial orders now get their derivatives as well, from the same single perturbation.

• Line targets could not use the model state at all. line_key::get_value had an inverted check (introduced in b5d0abb), so model_state_vecFromBands and abs_bandsFromModelState failed for every line parameter and line-shape target.

• Stale finalized flag. Adding targets after jac_targetsFinalize, or after jac_targetsInit, left the flag set, and x_size() returned wrapped-around sizes (e.g. 2^64 − 1). The new Targets::is_finalized() also catches targets appended from Python, and the covariance and OEM code now use it.

• Overlapping targets with transformations (relative, log, …):

  • the inverse Jacobian was applied once per group member (e.g. scaled by orig³ for the magnetic field);
  • jac_targetsAddOverlappingWindField moved from a copy and then reused it, so one component lost its transformation;
  • toggling after the overlap was added only changed the original.

  Transformations on overlap groups are now rejected. Only the original target updates the shared state and Jacobian columns.

• XML serialization of transformed targets. Only some atmospheric functors were registered. Log-relative and relative-humidity targets, any transformed surface or subsurface target, and every sensor or error polynomial target failed to save. All are now registered. The MK05 saturation-pressure operator is now a serializable functor, so water_equivalent_pressure_operator itself can also be saved.

• Registered function pointers in functional XML IO were matched by the address where they were stored, so they could never be written. Nothing used this path yet.

• RH targets with nonnegative=1 kept a non-zero Jacobian where the state is clamped to zero.

Behaviour changes

• jac_targetsFinalize now:
  • rejects duplicate error targets;
  • rejects line targets that are not in abs_bands;
  • requires overlapping wind or magnetic field targets to be magnitude fields (atm_fieldAbsolute*Field).
• Sensor frequency polynomial offsets now use the grid mapped to [−1, 1] as abscissa, so every coefficient is a shift in Hz (c₀ at the centre, c₀ ± c₁ at the edges). Previously coefficient k scaled with f^k (~1e11 per order).
  • Polyorder 0 is unchanged, and so is the XML format.
  • A polyorder ≥ the number of grid points is now rejected.
• Error targets now print their perturbation d like the other target types. The duplicate-sensor-target error no longer carries its own note.

Cleanup

• Targets::for_each_list replaces six copies of most per-type loops, including those in finalize.
• The four m_jactargets_*toggle.cc files are merged into one m_jactargets_toggle.cc with a single templated helper.
• The relative/log/log-relative functors use one call operator per signature, constrained on a transformable_field concept, and one templated make_*fit each. Before, there was one per field type.
• One std::formatter for all target types replaces six identical ones.
• Copy-pasted error messages are fixed, and the unused concepts target_comparable and valid_target and the redundant Jacobian::is_wind/is_mag are removed.

Tests

New:

• tests/core/jac/:
  • jac_targets_finalize.py
  • overlapping_targets.py
  • targets_xml_round_trip.py
  • line_targets_model_state.py
  • jac_targets_validation.py
  • rh_nonnegative_jacobian.py
• tests/core/sensor/sensor_poly_offset_jacobian.py
• cpp.fast.xml_functional

Extended: cpp.fast.covariance.workspace. The line targets in tests/core/jac/model_state_atm.py are re-enabled.

Notes for review

• The branch was written against the old ARTS_USER_ERROR_IF macro and rebased onto the change that removed it. The last commit converts the remaining uses, so the intermediate commits do not compile on their own. Squash-merging avoids that.
• The commits are ordered bug fixes, then validation, then refactors, and each has a detailed message.

🤖 Generated with Claude Code