test: registerRefImMeta is exercised against the function the database has

[RussLaher]

The gap this closes

rapid_db.register_refimmeta sends twenty-four values positionally into a
stored function declared by rapid_systems migration 128. The stub tier asserts
the argument order against the method's own signature — which proves the call
is self-consistent, not that it matches the function the database actually
has. Nothing in this repository contains that function, so nothing outside the
contract tier could.

That distinction is not hypothetical here. products.py records the same class
of defect reaching production: refimcatalogs.cattype is a smallint, the
registrar sent the strings "sextractor"/"photutils", PostgreSQL refused
every reference-image registration at its first catalogue, and the unit suite
stayed green throughout because its fake database accepts whatever it is
handed. A twenty-four argument positional call is that shape with more places
to be wrong.

This is the follow-up named in Caltech-IPAC/rapid#51's merge commit as the
one thing that change left unverified end-to-end.

What it asserts

Five tests in pipeline/contract/test_refimmeta_registration.py:

1. Every argument lands in the column it was meant for — column by column, not "a row exists", because a row-count check passes on a call whose arguments were rotated by one.
2. A second registration updates rather than inserting — the replay property the registration consumer depends on when a watermark write does not commit.
3. Metadata for an unregistered reference image is refused — through the wrapper, so the refusal is seen as exit_code 67, which is what products._check turns into RegistrationFailed.
4. A NULL measurement is refused by the column — the database's half of the registrar's own refusal to build a half-known row.
5. A refused call leaves the transaction aborted — see below.

Every value is distinct, and that is load-bearing. Two columns sharing a
value would make a swap between them invisible, which is exactly what an
argument-order test exists to catch. The real columns take binary32-exact
values (.25/.5/.75) so the round-trip comparison is exact and this suite
never fails on float formatting.

Test 5 is a finding, not a restatement

Writing these surfaced that register_refimmeta reports failure through
exit_code but leaves the transaction aborted — the next statement on that
connection raises InFailedSqlTransaction whatever it is.

The registration path is safe from this only because products._check raises
immediately after every call and the consumer's error path rolls back. A future
caller that "handled" exit_code 67 by carrying on would meet a wall of
errors naming nothing to do with reference images. It is asserted here so the
constraint is stated rather than rediscovered, and the tests that expect a
refusal take a SAVEPOINT for the same reason.

Proven able to fail

A test observed only passing proves nothing, so both were mutation-checked
against the wrapper, each mutation reverted:

Mutation	Caught as
swap nsxcatsources/npucatsources (same type — only the value check can catch it)	nsxcatsources expected 1234 got 1100; npucatsources expected 1100 got 1234
cast eight real parameters as integer	clmean expected 1.25 got 1.0; … cov5percent expected 87.25 got 87.0

Fixture

fixture.make_refimage is new and deliberately lighter than
_diffimage_parents: refimages has no FK to exposures or l2files, so a
caller needing only a reference image should not mint both to get one. fid is
read from the seeded filters rather than invented, and version is max+1
because refimagespk is UNIQUE on (field, fid, ppid, version) — a fixed
version would make the second call in a suite collide with the first, and the
failure would look like the defect under test.

Verification

• Contract tier: 744 passed, 15 skipped against PostgreSQL 18 + Q3C built from the full stream at rapid_systems main (e1896658, the commit carrying 128)
• Stub tier: 2677 PASS
• No production code changed — two files, both test-side.

What this still does not cover

The registrar's own assembly of those twenty-four arguments
(products._refimmeta_measurements and its REFIMMETA_MEASUREMENTS table) is
asserted in the stub tier against the mapping it declares. The two tiers
together cover the chain from provenance key to database column; neither covers
it alone. A full end-to-end register_reference_image against a real database
would need an attempt, a terminal record and published products, and is a
larger piece of work than this one.