test: reach 100% coverage

[Lancetnik]

Why the 98% was not what it looked like

[tool.coverage.report].exclude_also contained unanchored regexes. Coverage matches them against every line, and when a matched line belongs to a multi-line statement it drops the whole statement — for a def, that means the signature and the body.

'\.\.\.' therefore matched annotations such as tuple[Any, ...] and silently removed the core of the library from measurement:

excluded because of	what got dropped
params: tuple[OptionItem, ...],	all of CallModel.__init__
*args: tuple[Any, ...],	all of CallModel._solve
*args: tuple[Any, ...],	all of CallModel.solve
*args: tuple[Any, ...],	all of CallModel.asolve

That is also why Provider.merge showed up as uncovered while both of its call sites did not — the call sites were inside solve/asolve.

"from .*" and "import .*" were unanchored too, and swallowed ) from err continuation lines (so e.g. the msgspec raise ValidationError(...) from er statement was excluded).

What changed

Coverage config — anchor the stub-body patterns (^\s*pass$, (^|:)\s*\.\.\.$), replace the unanchored import patterns with if TYPE_CHECKING:, and drop the four patterns that matched nothing in this repo (self.logger, logger\..*, lambda: None, raise ValueError). Statements under measurement go from 1966 → 2846.

Tests for the real gap:

• Provider.merge — via a call-level dependency_provider= on solve/asolve (tests/test_provider.py)
• ExceptionGroup unwrapping for async field=True custom fields (tests/library/test_custom.py)
• evaluate_forwardref, eval_type_backport, is_backport_fixable_error (tests/test_compat.py)
• ValidationError.__str__ for both keyword and positional incoming options (tests/test_exceptions.py)
• SerializerProto.encode and Serializer.get_aliases defaults, plus a minimal custom serializer end to end (tests/library/test_serializer.py)
• msgspec use_fastdepends_errors=False and msgspec.field(name=...) aliases (tests/serializers/msgspec/test_serializer.py)
• pydantic non-class response types and the pre-2.11 model_fields fallback (tests/serializers/pydantic/)
• inject() with a prebuilt CallModel — the path integrations such as FastStream use (tests/test_inject.py)

The JSON backend branches, without new CI jobs — fast_depends.pydantic._compat picks orjson/ujson/stdlib at import time, and pydantic v1 routes dump_json through whichever won (v2 uses pydantic_core.to_json). Rather than a CI run per backend, tests/serializers/pydantic/test_json_backend.py re-executes the real module from disk against stubbed sys.modules entries. Coverage keys line data off the file, so the real branches are measured, and nothing is imported into the live module tree.

The v1 encode assertions are now separator-agnostic — the three backends differ only in separator whitespace, so a contributor who happens to have orjson installed locally would otherwise get spurious failures.

Source — two small changes:

• use.py: drop if SerializerCls is None: immediately after SerializerCls = None; the guard could never be false, so its second branch was unreachable.
• core/model.py: # pragma: no branch on for ex in exgr.exceptions: — the loop always raises on the first item and never exits normally.

CI — one line: coverage report in coverage-combine runs with --fail-under=100 so this cannot silently regress (plus if: always() so the html report is still uploaded when the gate fails). No new jobs; coverage stays cumulative across the existing ones.

Verification

Ran the CI matrix locally — py3.12+pydantic-v2, py3.12+pydantic-v1, py3.12+msgspec, py3.12 no-serializer, py3.10+pydantic-v2, py3.13+pydantic-v2 — all green, combined to 100%:

TOTAL   2846   0   320   0   100%

ruff check, mypy and codespell are unchanged from main (each still reports the same pre-existing findings; none are introduced or fixed here).

🤖 Generated with Claude Code