feat(chemistry): serve legacy water chemistry over REST - #870

Adds `result_kind` (major, minor, radionuclide, field) to the chemistry
results response.

Whether a result was read at the wellhead or by a laboratory is a
distinction an owner-facing report has to draw, and the legacy tables are
the only record of it -- the refactored `parameter_type` it used to come
from is not populated for this data. The view keeps that provenance only
in its text id prefix, so the prefix is translated here into a field a
client can read instead of every client learning to parse an id.

Refs BDMS-1189

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

GET /thing/water-well returned 500 for any page containing a thing with no
current location: `current_location` was a required field, so the GeoJSON
validator was handed None, reached for `__table__` on it, and raised
AttributeError. One unlocated well made every well on its page unreadable.
The dev database has 49 of them, so the listing failed at any page size
that reached one.

A thing is associated with a location over an effective period, and that
period can be closed or never opened, so having no current location is a
state the schema has to be able to say. The field is now optional and the
validator hands None back for the annotation to resolve.

Found while pointing the chemistry report at this endpoint; the bug is
older than that and affects every consumer of the wells listing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Working on the legacy chemistry REST routes means having legacy chemistry to
read, and until now the only way to get it into ocotilloapi_test was a SQL Server
connection or hand-written rows. Tests build their own single records inline,
which is right for unit tests but leaves nothing to exercise the normalized
chemistry views, the LIMS ingestion path, or a list endpoint against realistic
analyte, unit and detection-limit distributions.

This copies a bounded subset out of a local clone of another database (default
ocotillo_prod) into ocotilloapi_test, walking the dependency closure:

    thing -> location + location_thing_association
          -> NMA_Chemistry_SampleInfo
               -> NMA_MajorChemistry
               -> NMA_MinorTraceChemistry

The association rows matter as much as the locations. thing.nma_pk_location is a
legacy audit column, not the live link: the model reaches a location through
location_thing_association (Thing.location_associations), so seeding a thing
without association rows produces exactly the location-less well that bff7faa
had to stop 500ing on.

Primary keys are deliberately not preserved. The target already holds unrelated
rows at low ids, so copying source ids verbatim would silently reparent sample
infos onto pre-existing test things wherever the ranges overlap. Rows are
inserted without an id and children are repointed at the new parent id.

Re-runs reconcile on the legacy natural keys rather than on ids: a sample info
whose nma_SamplePtID is already present is skipped, and location/thing are reused
via nma_pk_location/nma_pk_welldata. So a second run picks up the next unseeded
batch instead of duplicating the last one.

Candidates must have both a major and a minor/trace row, so every seeded sample
exercises both tables. thing.thing_type is a NOT NULL lexicon FK, so a thing
whose type the target lexicon lacks disqualifies its sample infos rather than
being patched; nullable lexicon-backed columns are nulled with a count reported.
Column sets are intersected per run, so schema drift between the two databases
degrades to a printed skip list instead of a crash. thing.search_vector is
skipped as trigger-maintained, and location.point round-trips as EWKT because
pg8000 carries no geometry codec.

The target name must contain 'test' unless --force is passed.

The seed is transient by design of the test suite: the session-scoped autouse
fixture in tests/conftest.py drops and re-migrates the schema, so any pytest run
wipes it and the script has to be re-run afterwards. That also means it cannot
perturb the suite -- pytest always starts from a clean schema. Documented in the
module docstring.

Verified by seeding 65 sample infos across two runs (764 major and 1223
minor/trace rows, 22 and 79 distinct analytes, 437 censored values): the second
run reported 60 already-seeded candidates skipped, natural keys stayed unique
(65/65 sample infos, 764/764 major GlobalIDs), no chemistry row was orphaned, all
60 chemistry things resolved a location through the association table, and every
copied location kept its geometry.

Depends on a local database clone, so it is a dev-box seeder and does not run in
CI.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>