refactor(ogc): close over request state, retiring ogc.context

[thodson-usgs]

What

Retires the ambient trio — ogc/context.py's base-URL, dialect, and row-cap ContextVars — in favor of explicit binding, and drops the copy_context() snapshot from FanOut.

Before this change the same three values traveled by three mechanisms at once: partial-bound finalizer arguments, ambient ContextVars, and .get() fallbacks deep inside request construction and shaping. No signature stated they were required, and the service-neutral FanOut executor snapshotted an entire execution context at construction (with a ten-line comment explaining which adapter ambients it must not drop) solely so a resume fired after the originating with-blocks exited would rebuild chunks against the right API.

How

• _construct_api_requests / _construct_cql_request take base_url (and dialect) as explicit parameters. get_ogc_data binds them with functools.partial — the pattern its finalizer already used — into both the plan's build_request and the per-chunk fetch, so planning-time sizing and a later exc.call.resume() rebuild every chunk against the values the call was created with.
• row_cap is a real parameter of _paginate/_walk_pages, forwarded to the paginate parameter transport already had, instead of an ambient read inside a compatibility wrapper.
• _finalize_ogc / _deal_with_empty require base_url; the empty-frame schema lookup can no longer silently target an ambient default.
• FanOut loses self._ctx and _resume_in_context: transport carries no adapter state (ADR 0006), and the progress reporter reaches the worker thread through the portal's normal calling-context copy.
• waterdata.get_cql states its target explicitly instead of importing a private ambient and entering its scope.
• ogc/context.py is deleted; ADR 0002's prose and the architecture overview are updated to match.

Why

Interfaces now state their preconditions (the base_url-omitted footgun documented on get_ogc_data is caught by the signature everywhere, not just at the one required call site), the resume-after-return invariant is structural rather than guarded by a snapshot and prose, and the one adapter-shaped concern inside service-neutral transport is gone.

Testing

• Full offline suite: 780 passed.
• mypy --strict, ruff check/format, xenon, complexipy, and lint-imports (7 contracts) all pass.
• The captured-context regression test is rewritten as a creation-time-binding test; the direct request-construction unit tests bind the Water Data target with the same partial pattern the engine uses.

🤖 Generated with Claude Code

[thodson-usgs]

Health metrics vs main

Measured locally with the same tooling CI runs (pyscn 1.29.0, radon, xenon, complexipy, import-linter), each branch in its own worktree so import resolution is comparable — ProjectRoot verified to equal the checkout in every run, per the sanity check in code-health.yml.

pyscn composite: 82/100 (B) — unchanged, and all seven sub-scores are identical to main (Complexity 95, Dead code 100, Duplication 70, CBO 100, LCOM 100, Dependencies 80, Architecture 87).

Underneath the banded scores:

	main	this PR
Modules	57	56
Resolved dependency edges	140	135
Max dependency depth	8	8
Technical-debt estimate	162 h	160 h
Functions	363	361
High-risk entries	19	19
Package LOC (non-blank)	13,090	13,079

Per-function movement (radon cyclomatic):

• _finalize_ogc 8 → 7, _deal_with_empty 4 → 3 — the ambient .get() fallbacks were branches.
• _construct_api_requests 5 → 6 — the dialect is None default is a new branch; that is the cost of making the parameter explicit.
• Removed: FanOut._resume_in_context, ogc/context.py.

Maintainability index (radon): ogc/requests.py 48.3 → 49.6, ogc/policy.py 56.1 → 57.2, ogc/engine.py 61.7 → 61.9, waterdata/cql.py 83.4 → 84.1; ogc/shaping.py 45.1 → 44.8 and transport/fanout.py 27.7 → 27.5 are the two small regressions. All stay rank A.

One number that looks like a regression but isn't: duplication reads 8.83% → 8.90%. Clone groups (5), clone pairs (50), and total clones (25) are identical — no new duplication. The package got smaller, so the same clones are a larger fraction of it.

Merge gates (xenon --max-absolute C --max-modules B --max-average A, complexipy, lint-imports, mypy --strict): pass, as on main.