farallon_bird-mammal: cruise_key is NULL on all 66,272 obs rows

[bbest]

In release v2026.08.04, every obs row for farallon_bird-mammal has a
NULL cruise_key:

SELECT dataset_key, count(*) AS obs, count(cruise_key) AS with_cruise,
       count(DISTINCT cruise_key) AS n_cruises
FROM read_parquet('.../v2026.08.04/parquet/obs.parquet')
WHERE dataset_key IN ('farallon_bird-mammal','cce-lter_zoodb','swfsc_ichthyo')
GROUP BY 1;

dataset_key	obs	with_cruise	n_cruises
cce-lter_zoodb	18,276	17,022	35
farallon_bird-mammal	66,272	0	0
swfsc_ichthyo	459,286	459,286	671

Why it matters

cruise_key is denormalized onto obs specifically so consumers can roll up by
survey without a join. Any consumer doing count(DISTINCT cruise_key) reports
zero for this dataset — not "unknown", but a confident zero.

It's already visible to users: the station portal's coverage card reads
"0 surveys · 760 obs" for Seabirds & Marine Mammals at station 090.0 055.0,
and does so for all 101 of its stations. Other datasets have only 1–3
stragglers each. Screenshot in CalCOFI/db-viz-station#10.

The observations themselves look fine — 760 obs across 457 samples at that
station, 1987-05-05 → 2018-04-10 — so this is a missing FK stamp, not missing
data.

Suggested fix

In the bird/mammal ingest's "Emit Core Tables" section, resolve each transect to
its cruise and pass cruise_key through to append_obs() the way the other
ingests do. If these transects genuinely have no CalCOFI cruise association,
that's worth stating explicitly in the dataset metadata so consumers can
distinguish "not applicable" from "not populated" — a NULL FK currently reads as
the latter.

Worth checking whether sample.cruise_key has the same gap.