Extract _DecoderResolver for Seam B

[lesnik512]

Deepening candidate #3 from the architecture review.

Problem

The decoder seam (Seam B) is real — pydantic + msgspec adapters — but the logic for using it was copy-pasted across four send/send_with_response methods on the two clients (client.py:185, 213, 1200, 1228): walk _decoders, raise MissingDecoderError(registered_names=…) pre-flight if none claims the model, then decoder.decode(...) wrapping failures as DecodeError. Plus _dispatch_decoder defined twice.

Change

One _DecoderResolver (+ generic _BoundDecoder) in a new decoders/_resolver.py owns it:

• resolve(model) -> _BoundDecoder[T] — first can_decode wins; none → MissingDecoderError (pre-flight, before the HTTP call).
• _BoundDecoder.decode(response) -> T — wraps except Exception as DecodeError, with the decoder + model sealed together so a mismatch is unrepresentable.

The four sites collapse to resolve → dispatch → bound.decode. Both clients keep self._decoders (tested attribute) and gain self._decoder_resolver; both _dispatch_decoder methods removed. Fully synchronous → one resolver serves both clients. Behaviour byte-identical.

Tests

• Existing decoder/client suites stay green (parity net).
• New tests/test_decoder_resolver.py drives resolve/bound.decode directly: resolution, MissingDecoderError with registered_names + ordering, first-match-wins, DecodeError wrapping.

Verification

• just lint clean (ruff + ty), just test → 723 passed, 100% coverage
• grep -rE 'httpx2\._' src/httpware/ clean

Design bundle: planning/changes/2026-06-23.02-decoder-resolver-extraction/. Promoted into architecture/decoders.md.

🤖 Generated with Claude Code