fix(params): load flat-model parameter values correctly + save calibrated model (#114)

[FinbarArgus]

Fixes #114 (DEFCON 1): uploading a circulatory_autogen flat CellML loaded
the wrong parameter values, so simulations/plots didn't match the file — for
calibrated and non-calibrated models.

Root cause

A flat model renames its constants. params_for_id aortic_root/C lives in the
model as parameters/C_aortic_root; global/E_lv_A as parameters_global/E_lv_A.
The loader only looked up the direct vessel/param qname, didn't find it, so the
slider initial_value came back None — and the UI substituted the range
midpoint. Wrong values in → wrong sim out.

Fix

params_for_id.resolve_model_qname() resolves the value by CA's own name mapping
when the direct qname is absent: it computes the param_names_for_gen name (param
for the global vessel, else param_vessel, exactly per PrimitiveParsers) and
looks it up by bare variable name — preferring the flat model's parameters*
component and only when unambiguous. Direct vessel/param still wins first, so
non-flat models (Lotka-Volterra) are unaffected.

Verified: the 4 params of the 3compartment flat model now resolve to their real
file values (were all None/midpoint before).

Save the calibrated model (the reported test path)

Per "make sure the cellml model is saved when finishing calibration":

• calibrated_model.calibrated_cellml() bakes best-fit values into the flat
  model's initial_value attributes — pure-XML, component-scoped substitution
  (the exact inverse of the loader), so the rest of the document is byte-identical.
• calibration_runner writes <prefix>_calibrated.cellml on finish; the path is
  threaded into the job status and served at
  GET /api/calibration/{job}/calibrated_model.
• The Calibration panel shows a Download calibrated model link when a run is
  done.

Tests

• Unit: gen-name resolution (incl. ambiguity + parameters* preference);
  calibrated-model substitution, "only targeted lines change", round-trip, and the
  download route (404s + happy path).
• Integration (the exact broken scenario): run a real GA calibration →
  the saved calibrated model has the best-fit values → re-upload it + params_for_id
  ("new window") → the slider initial values equal the best fit.
• Backend 350 unit / 378 incl. integration; frontend 297; yarn build clean.

🤖 Generated with Claude Code