docs(planning): update Seam B for multi-decoder routing

Author: lesnik512

planning/engineering.md

@@ -4,7 +4,7 @@ This doc is the single distilled reference for `httpware` design rationale, prot
 
 ## 1. Project intent
 
-`httpware` is a thin opinionated wrapper around `httpx2`. It re-exports `httpx2.Request` and `httpx2.Response` as the public request/response surface and adds three things on top: typed response decoding (via a `ResponseDecoder` protocol; pydantic and msgspec are both opt-in extras as of 0.3.0), a middleware chain composed at client construction, and a status-keyed exception tree raised automatically on 4xx and 5xx. `AsyncClient(decoder=None)` defaults to constructing a `PydanticDecoder` and so requires the `pydantic` extra; callers can supply an explicit `decoder=` argument to escape the default. As of 0.4.0, the package ships a small resilience suite under `httpware.middleware.resilience` — a `Retry` middleware with a Finagle-style `RetryBudget`, plus a `Bulkhead` concurrency limiter — composed via the standard middleware chain. As of 0.5.0, `AsyncClient.stream()` provides a context-manager API for chunked response bodies; it bypasses the middleware chain by design (see planning/archive/specs/2026-06-05-streaming-design.md). As of 0.6.0, `Retry` and `Bulkhead` emit operational events via stdlib `logging` records (`httpware.retry` / `httpware.bulkhead` loggers) and — when `opentelemetry-api` is installed — OpenTelemetry span events on the active span. As of 0.7.0, the first-cut user-docs surface is live at <https://httpware.readthedocs.io/> (Middleware, Resilience, Errors, Testing guides) and Epic 3 is closed.
+`httpware` is a thin opinionated wrapper around `httpx2`. It re-exports `httpx2.Request` and `httpx2.Response` as the public request/response surface and adds three things on top: typed response decoding (via a `ResponseDecoder` protocol; pydantic and msgspec are both opt-in extras as of 0.3.0), a middleware chain composed at client construction, and a status-keyed exception tree raised automatically on 4xx and 5xx. As of 0.9.0, both clients take `decoders: Sequence[ResponseDecoder] | None = None` (a *list*, not a single instance) and dispatch via each decoder's `can_decode(model)` predicate; the default resolves against installed extras (pydantic-first when both present) and `AsyncClient()` / `Client()` no longer raise on missing extras. A new `MissingDecoderError` (sibling of `DecodeError` under `ClientError`) fires before the HTTP call when `response_model=` is set but no registered decoder claims the model. As of 0.4.0, the package ships a small resilience suite under `httpware.middleware.resilience` — a `Retry` middleware with a Finagle-style `RetryBudget`, plus a `Bulkhead` concurrency limiter — composed via the standard middleware chain. As of 0.5.0, `AsyncClient.stream()` provides a context-manager API for chunked response bodies; it bypasses the middleware chain by design (see planning/archive/specs/2026-06-05-streaming-design.md). As of 0.6.0, `Retry` and `Bulkhead` emit operational events via stdlib `logging` records (`httpware.retry` / `httpware.bulkhead` loggers) and — when `opentelemetry-api` is installed — OpenTelemetry span events on the active span. As of 0.7.0, the first-cut user-docs surface is live at <https://httpware.readthedocs.io/> (Middleware, Resilience, Errors, Testing guides) and Epic 3 is closed.
 
 As of 0.8.0 the async middleware surface uses the `Async*`/`async_*` prefix (aligning with httpx2's convention); the `attempt_timeout=` kwarg was removed from `AsyncRetry` in the same release — see `planning/specs/2026-06-07-sync-client-design.md` for the rationale.
 
@@ -36,11 +36,15 @@ The 0.1.0 seams numbered 1 (Middleware↔Transport) and 4 (Transport↔httpx2) h
 - **Contract:** the middleware chain is composed once at client construction and frozen for the client's lifetime. Both worlds follow the same contract; the only difference is the per-world type: `AsyncClient` composes `AsyncMiddleware` via `compose_async` (the continuation type is `AsyncNext`), and `Client` composes `Middleware` via `compose` (the continuation type is `Next`). Both `compose` and `compose_async` live in `src/httpware/middleware/chain.py`. The chain bottom (the "terminal") is internal: it calls `self._httpx2_client.send(request)`, maps `httpx2` errors to `httpware` errors, and raises a `StatusError` subclass on 4xx/5xx. Same lifecycle rules in both worlds.
 - **Rule:** mutating the chain after construction is not supported. Per-request behavior goes through `httpx2.Request.extensions` or through `extensions=` kwargs at call sites.
 
-### Seam B: `Client`/`AsyncClient` ↔ `ResponseDecoder`
+### Seam B: `Client`/`AsyncClient` ↔ `ResponseDecoder` list
 
 - **Where:** `src/httpware/client.py` ↔ `src/httpware/decoders/`.
-- **Contract:** the decoder is invoked when the caller passes `response_model=`. The protocol is `decode(content: bytes, model: type[T]) -> T`. Any exception raised by `decode` is wrapped by the call sites in `client.py` — `Client.send` / `AsyncClient.send` (when `response_model=` is set) and `Client.send_with_response` / `AsyncClient.send_with_response` — into `httpware.DecodeError` (a `ClientError` subclass carrying `response`, `model`, `original`). Decoder implementers do not need to raise `DecodeError` directly.
-- **Rule:** the decoder must operate on raw bytes in a single parse pass. Two-pass decoding (`json.loads` then `validate_python`) is rejected: a single bytes-in / typed-object-out pass avoids the redundant intermediate `dict` allocation and parses faster. The Pydantic adapter implements this as `TypeAdapter(model).validate_json(content)`, with the `TypeAdapter` itself memoized via `@functools.lru_cache(maxsize=1024)` on a module-level `_get_adapter(model)` factory (the adapter is the expensive part to build). The msgspec adapter implements it as `msgspec.json.decode(content, type=model)`.
+- **Contract:** the client holds `_decoders: tuple[ResponseDecoder, ...]` composed at `__init__` and frozen for the client's lifetime. The Protocol exposes two methods:
+    - `can_decode(model: type) -> bool` — predicate used at send-time to walk `_decoders` and pick the first claiming decoder (`_dispatch_decoder` on both classes). Built-in decoders claim broadly (pydantic via `TypeAdapter(model)` probe, msgspec via `msgspec.inspect.type_info(model)` + `CustomType` filter); list ordering decides ambiguous shared shapes (dataclass, primitive, generic). Native types of another library MUST be rejected.
+    - `decode(content: bytes, model: type[T]) -> T` — the decode itself. Any exception is wrapped by `Client.send` / `AsyncClient.send` (when `response_model=` is set) and `Client.send_with_response` / `AsyncClient.send_with_response` into `httpware.DecodeError` (a `ClientError` subclass carrying `response`, `model`, `original`). Decoder implementers do not need to raise `DecodeError` directly.
+- **Pre-flight check:** when `response_model=` is set and no decoder claims it, `send` / `send_with_response` raise `MissingDecoderError(model=..., registered_names=...)` BEFORE the HTTP call. Distinct from `DecodeError` (which means the decoder ran and the payload was malformed); distinct corrective actions (install an extra or pass `decoders=[...]`).
+- **Default list:** `decoders=None` resolves via `client.py:_build_default_decoders()` against installed extras — pydantic-first when both are present, either-only when only one is installed, empty tuple when neither. `AsyncClient()` / `Client()` never raise on missing extras; failure surfaces only at the first `response_model=` use site.
+- **Rule:** the decoder must operate on raw bytes in a single parse pass. Two-pass decoding (`json.loads` then `validate_python`) is rejected: a single bytes-in / typed-object-out pass avoids the redundant intermediate `dict` allocation and parses faster. The Pydantic adapter implements this as `TypeAdapter(model).validate_json(content)`, with the `TypeAdapter` itself memoized via `@functools.lru_cache(maxsize=1024)` on a module-level `_get_adapter(model)` factory; the msgspec adapter mirrors the pattern with a cached `msgspec.json.Decoder(model)`.
 
 ### Seam C: `httpware ↔ optional extras`
 
@@ -118,7 +122,7 @@ Each extra's code lives in a single dedicated module (`decoders/pydantic.py`, `d
 
 New extras are added at the same time as the code that uses them — never preemptively. (An `otel` extra existed pre-0.4 but was removed once we noticed it was advertising functionality that didn't exist. 0.6.0 reintroduces it paired with the code that uses it — `Retry` and `Bulkhead` add events to the active OpenTelemetry span via `trace.get_current_span().add_event(...)`.)
 
-Caller-facing pattern: consumers select the implementation by passing it explicitly, e.g., `AsyncClient(decoder=PydanticDecoder())`. There is no auto-detection or implicit registry.
+Caller-facing pattern: as of 0.9.0, `AsyncClient()` / `Client()` default `decoders=None` resolves via `_build_default_decoders()` against installed extras (pydantic-first when both are present; empty tuple when neither). Consumers override by passing `decoders=[...]` explicitly; `decoders=[]` is honored as an opt-out. The auto-resolution is a snapshot of `import_checker.is_<extra>_installed` flags at `__init__` time — there is no runtime re-detection or implicit registry beyond the two built-in decoders.
 
 ## 8. Remaining roadmap