test(obs_data): prove per-data_item operation_kwargs actually take effect

[FinbarArgus]

Follow-up to #113. Its tests proved the kwarg value is carried (signature →
schema, UI renders/persists, JSON round-trip) but never that it has an effect —
so the feature could break end-to-end with a fully green suite. This adds the three
missing links.

1. The value reaches CA's input file

test_obs_data_upload.py — operation_kwargs survive the upload into the stored
obs_data.json that CA actually reads (asserted on both the response payload and
the file on disk). parse_obs_data currently passes data_items through verbatim;
a whitelist-style refactor would silently drop the field and every existing
kwargs test would still pass.

2. The value reaches the operation call

test_local_sensitivity.py — _evaluate_features splats operation_kwargs into
func(*operands, **kwargs). Parametrised so an ignored kwarg fails loudly
(factor 1/2/4 → 3/6/12; if kwargs were dropped all three collapse to 3). Also
covers the no-kwargs default and the documented substitution of a string kwarg
naming an earlier observable's feature — two call sites (lines 222, 355) that had
no coverage at all.

3. End-to-end: the value changes the result

test_user_funcs.py (@pytest.mark.integration) — on a CUFLynx-authored custom op
(#104), against the real CA:

1. the op's kwarg is discoverable via operation_kwargs_schema (and the reserved
   series_output flag is not offered as a tunable),
2. CA parses our obs_data.json and keeps operation_kwargs through
   parse_obs_data_json → process_obs_info,
3. invoking the registered op the way CA does gives a different value per kwarg
   (factor 2 → 6, factor 4 → 12, default → 3).

Skips cleanly when CA isn't importable or predates the external-path support (#303).

Tests

Backend 330 unit passed, 360 incl. integration. Tests-only — no production
code changed.

🤖 Generated with Claude Code