fix(validation): bundle schemas before wheel builds

[bokelley]

Summary

• run schema bundling before release-package builds so published wheels/sdists include adcp/_schemas
• run schema bundling in the CI wheel smoke build so tested wheels match release packaging
• surface missing SDK-pinned schema bundles at warning level while keeping explicit unsupported version= misses quiet

Fixes #897

Validation

• expert review: DX/release packaging pass found the direct CI wheel-build gap; addressed in this PR
• expert review: code review pass found explicit-version warning noise; addressed with regression tests
• uv run --extra dev --group dev pytest tests/test_schema_loader_per_version.py tests/test_legacy_schema_bundle_v2_5.py tests/test_schema_validation.py tests/test_validation_version.py
• uv run --extra dev --group dev pytest tests/ -v --cov=src/adcp --cov-report=term-missing
• uv run --extra dev --group dev ruff check src/ tests/test_schema_loader_per_version.py
• uv run --extra dev --group dev mypy src/adcp/
• uv run --extra dev --group dev mypy --strict tests/type_checks/
• uv run --extra dev --group dev python scripts/check_type_ignore_contract.py
• uv run --extra dev --group dev python scripts/bundle_schemas.py && uv run --extra dev --group dev --with build python -m build --wheel --outdir dist/
• verified built wheel contains 3468 adcp/_schemas/*.json entries

[aao-ipr-bot]

LGTM. Fix is the right shape — bundle before build at both CI smoke and release plumbing, and warn (not debug) when the SDK pin can't find its own bundle.

Things I checked

• _ensure_state discrimination at src/adcp/validation/schema_loader.py:189-197: log_missing = logger.warning if version is None else logger.debug is correct. SDK-pinned miss = packaging bug worth surfacing; explicit version= miss = caller asked for a bundle we don't ship, stays quiet. Fail-closed beats fail-open here.
• One-shot warning behavior. _state_misses.add(bundle_key) at L196 plus the early returns at L179 and L187 mean the warning fires exactly once per bundle_key per process. No log spam in long-lived workers.
• Both new tests at tests/test_schema_loader_per_version.py:194-223 wrap in _reset_for_tests() before and after, so _state_misses is clean across the assertion window. The monkeypatch on _resolve_schema_root exercises the branch as intended.
• Workflow gating in .github/workflows/release-please.yml:98-100: same release_created || workflow_dispatch+publish condition as the existing build step. Symmetric.
• CI smoke build at .github/workflows/ci.yml:171-175 now bundles before python -m build --wheel, so the smoke wheel matches the release wheel shape. The validation gap that hid #897 is closed.
• No public-API drift. schema_loader is reachable from adcp.validation, but the change is logger-level only.

Follow-ups (non-blocking — file as issues)

• Wheel-content assertion in CI. Three lines (python -m zipfile -l dist/*.whl | grep -q 'adcp/_schemas/.*\.json') on the smoke job would have caught #897 the first time, and is cheap insurance against any future regression in bundle_schemas.py, package-data globs, or workflow ordering. Notable that a PR landing schema-bundling into CI doesn't also assert the schemas ended up in the wheel.
• Move bundling into the build backend. Anyone running python -m build outside this workflow — local release, downstream rebuild, uv build, vendor repackaging — reproduces #897. A setuptools build_py subclass or an in-tree PEP 517 backend wrapper would make the wheel correct by construction. The current CI-step approach is a workaround at the wrong layer.
• warnings.warn vs logging.warning for the implicit-bundle miss. logger.warning is invisible to adopters who haven't configured the adcp.validation logger — the common case on a first-touch broken install. A once-per-process warnings.warn would surface in pytest and most app harnesses without logger config. Not blocking — warning is a strict improvement over debug either way.

Minor nits (non-blocking)

1. Negative test could be tighter. tests/test_schema_loader_per_version.py:215-223 asserts the specific string isn't in caplog.records. Asserting not any(r.levelno >= logging.WARNING for r in caplog.records) would also catch a regression that emits a different message at warning level on the explicit-version path.

Approving on the strength of the gating-condition symmetry plus the one-shot-warning behavior.