Glossary violations across the package: full inventory for planning #406

Scan of the whole package against CONTEXT.md as #405 leaves it, so the total
cost is visible before anything is scheduled. Grouped by what a fix actually
costs, not by severity.

Ground rules used, both from ADR 0013:

• Core terms (ours — chunk, page, source, fan-out, dialect, leaf)
  take one spelling everywhere, identifiers included. A second spelling is a defect.
• Domain terms (monitoring location, collection) are fixed for prose only.
  An adapter keeps its service's spelling at its public surface, and that is not
  a violation. So nwis.get_record(service="dv") and wqp's Station stay.

What that leaves is below. Nothing here is a behaviour bug.

1. Internal, mechanical — no public surface

Safe to do in one PR each, or one PR total.

• OgcDialect.cql2_services / date_only_services are keyed by collection
  (ogc/policy.py:45,48,65,66, read at ogc/requests.py:182,186). Already
  recorded under Known legacy names as the last two holdouts of the
  service→collection rename. Internal dataclass fields; renameable.
• service_url names a collection's items URL — ogc/requests.py:173,227.
  Local variables.
• Shared OGC machinery says site in prose. It serves both Water Data and
  NGWMN, so nothing there is about one service's spelling:
  ogc/shaping.py:93, ogc/planning.py:64,193,205, ogc/chunking.py:4,
  ogc/filters.py:60. Note planning.py uses site= / sites=[...] in
  illustrative URLs — Water Data's real parameter is
  monitoring_location_id, so those examples are stale as well as off-glossary.
• Water Data adapter prose says site where its own parameter is
  monitoring_location_id: waterdata/measurements.py:196,204,212,255,341,352,559,
  waterdata/cql.py:118, waterdata/nearest.py:269, waterdata/ratings.py:138.
  Excluded deliberately: "site visit" (measurements.py:4,49) is a
  field-practice term of art, not the place — leave it.
• batch used informally for a fan-out's set of chunks —
  exceptions.py:332,339, transport/fanout.py:520, waterdata/ratings.py:135,138.
  Lowest value on this list; listed for completeness.

2. The service= keyword on shared helpers — one decision, ~10 call sites

The largest single cluster, and the one worth deciding before the mechanical work.

Three shared entry points take a service argument:
ProgressReporter(service=) (progress.py:108), FanOut(service=)
(transport/fanout.py:255), and resolve_next_url(service=) (transport/links.py:44).

What callers actually pass:

progress.py:117 concedes it in its own comment: "The service/collection being
retrieved (e.g. daily, peaks)". It is the progress line's leading label and
the error message's subject — it is whatever names the thing being retrieved,
which is why five different kinds of name arrive there.

CONTEXT.md currently blesses this and should not. Known legacy names says
service "still means the external system in transport and progress, where
it labels a progress line. That usage is correct." The call sites say otherwise —
nwdc.py:387 passes the same string as both service= and adapter=. That
sanction needs to go whichever way the rename lands.

Options, roughly: rename the parameter to what it is (label? subject?) and
let each caller pass its own accurate name; or keep service= and fix only the
callers that can honestly supply a service. Both are internal — none of these
three is public API. Needs a call before the Tier 1 work, since it touches the
same files.

3. Public surface — needs an ADR 0012 deprecation cycle

• wqp.services_wqx3 / wqp.services_legacy (wqp.py:65,66) are public
  module-level lists holding WQP profiles (Result, Station,
  Activity) — collections by the glossary. Previously reviewed and
  deliberately left alone; on the list because it cannot be fixed without a
  deprecation.
• wqp.wqp_url(service) / wqp.wqx3_url(service) — both in __all__
  (wqp.py:688,696), parameter is a profile/collection.
• "<service default>" printed by show_configuration()
  (configuration.py:837). The glossary calls this an adapter default.
  User-visible output, so it is a compatibility question rather than a
  docstring edit. Raised in refactor: one word per concept in the configuration chain #400 and deliberately left there, since that PR's
  claim is that behaviour is unchanged.

4. Service vs adapter in the configuration modules

~11 sites where prose says service for what CONTEXT.md is emphatic is the
adapter ("The scope is the adapter, not the service and not the host"):
configuration.py:234,554,634,636,646,648, waterdata/configuration.py:6,34,47,60,
_configuration_core.py:28,225,393.

Not uniformly wrong — a base URL arguably is the service's even when the
setting is adapter-scoped, so this needs a judgement per site rather than a
sweep. #400 fixes a few in passing; the rest are open.

5. Settled — listed so they are not re-raised

• waterdata.get_cql(service=) — deprecated, removal 2027-08-09.
• waterdata.WATERDATA_SERVICES — permanent alias of WATERDATA_COLLECTIONS.
• waterdata.get_samples(service=) and SERVICES — names a resource, not a
  service or a collection; kept by decision, reasoning in Known legacy names.
• waterdata.get_codes(code_service=) — reproduces the Samples API's own word.
• utils.query — one request, not a query; frozen public path.
• dataretrieval.nwis service= throughout — ADR 0005 quarantine, frozen.
• NGWMN "provider" — that service's own vocabulary.
• ChunkInterrupted, ChunkedCall — permanent aliases.

Suggested order

Decide §2 first (it dictates edits in the same files as §1), then §1 as one or
two PRs, then §4 site by site. §3 is a separate deprecation-cycle decision with
no deadline pressure.

Depends on #405 for the rules this is scanned against.