Problem

Fixes #1136.

A Raw-typed field fed a plain decoded value through convert() (and therefore through yaml.decode/toml.decode, both of which parse first and call convert() on the result) raises a ValidationError that says the expected type is any:

>>> class Repro(msgspec.Struct):
...     tools: dict[str, msgspec.Raw]
>>> msgspec.yaml.decode(b"tools:\n  x: 1\n", type=Repro)
msgspec.ValidationError: Expected `any`, got `int` - at `$.tools[...]`

That's self-contradictory -- if any were really acceptable, there'd be no error at all.

Root cause

msgspec.Raw fields are marked with a typecode of 0 (see the comment at _core.c's type-node builder: "Raw is marked with a typecode of 0, nothing to do"). For json.decode/msgpack.decode, that's fine: the raw byte span for a Raw field is captured up front, before any type-checking against type->types happens, so a typecode-0 mismatch is never actually reached through those paths.

convert() has no equivalent interception -- it dispatches purely on the input Python object's type (convert_int, convert_str, convert_dict, ...), and each of those falls through to a generic "no match" error when none of the target's type->types bits are set. That generic path calls typenode_simple_repr() to build the human-readable "expected" string, and that function treats typecode 0 identically to real Any (bit MS_TYPE_ANY set), returning "any" for both. Real Any fields always set the MS_TYPE_ANY bit explicitly (state->types = MS_TYPE_ANY), so types == 0 is unambiguously the Raw marker -- but the repr function was conflating the two.

The one already-supported case (convert({"x": raw_instance}, type=Ex), where the input is already a Raw object) goes through a separate convert_raw dispatch keyed on the input's Python type, and correctly labels a rejected input as `raw` in the reverse direction. This PR makes the forward direction consistent with that existing wording.

What changed

• src/msgspec/_core.c: typenode_simple_repr() now returns "raw" for typecode 0 instead of "any", so the error reads Expected `raw`, got `int` -- accurate, and consistent with the wording already used for the reverse mismatch direction. Real Any-typed fields are unaffected (confirmed they always carry the explicit MS_TYPE_ANY bit, never falling into the types == 0 branch).
• tests/unit/test_convert.py: added a parametrized regression test (TestRaw.test_raw_mismatch_error_message) covering int/str/array/object inputs against a Raw-typed field.
• docs/supported-types.rst: added a note under the Raw section clarifying that Raw fields are only populated by json.decode/msgpack.decode (which can capture original encoded bytes before decoding further); convert() and anything built on it (yaml.decode, toml.decode) operates on already-parsed Python objects, so there's no encoded-byte buffer left to capture, and only an already-constructed Raw object is accepted as input.

Explicitly out of scope

The issue (filed as a question) could also be read as "should convert() support building a Raw from an arbitrary already-decoded value?" -- e.g. by re-serializing it to some byte format. I deliberately didn't attempt that here: Raw's whole documented purpose is "already-encoded bytes, so the encoder can skip re-encoding it," and a plain Python dict decoded from YAML isn't already encoded in any particular wire format, so picking one (JSON? msgpack?) to synthesize would be inventing new semantics rather than fixing a bug. If that's wanted, it feels like a separate design decision -- happy to hear if Raw is intended to be constructible via convert() at all before attempting it.

How it was verified

• Reproduced the exact scenario from msgspec.Raw appears not to work with msgspec.yaml.decode #1136 (yaml.decode with a dict[str, msgspec.Raw] field, five different YAML scalar/collection shapes) against unmodified main first, confirming the misleading Expected `any` message.
• After the fix, all five now raise Expected `raw`, got `<type>` as expected; the existing supported case (convert({"x": raw_instance}, type=Ex)) still round-trips correctly; a genuine Any-typed field mismatch is unaffected.
• just rebuild=1 test-all: 6414 passed, 113 skipped (0 failed). Ran the identical command against unmodified main for comparison: 6410 passed, 113 skipped -- the +4 delta is exactly the new parametrized regression test (int/str/array/object), zero regressions.
• just test-doc: 7 passed (doctests unaffected).
• just hooks-check (ruff lint, ruff format, codespell): all passed.
• just doc-build: Sphinx build succeeds cleanly with the new supported-types.rst note.