refactor(decoders): share memoizing get-or-build cache across adapters

architecture/decoders.md

@@ -12,4 +12,5 @@ Both clients take `decoders: Sequence[ResponseDecoder] | None = None` (a *list*,
     - `decode(content: bytes, model: type[T]) -> T` — the decode itself. Any exception is wrapped by `_BoundDecoder.decode` — the bound decoder `resolve` returns, sealing decoder + model together — into `httpware.DecodeError` (a `ClientError` subclass carrying `response`, `model`, `original`); this runs for `send` / `send_with_response` on both clients when `response_model=` is set. Decoder implementers do not need to raise `DecodeError` directly.
 - **Pre-flight check:** when `response_model=` is set and no decoder claims it, `_DecoderResolver.resolve` (called by `send` / `send_with_response` before `_dispatch`) raises `MissingDecoderError(model=..., registered_names=...)` BEFORE the HTTP call. `MissingDecoderError` is a sibling of `DecodeError` under `ClientError`, and is distinct from it: `DecodeError` means the decoder ran and the payload was malformed; the two have 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` cached per-instance on `PydanticDecoder._adapters: dict[type, TypeAdapter]` (populated lazily on first `_get_adapter()` call); the msgspec adapter mirrors the pattern with `MsgspecDecoder._msgspec_decoders: dict[type, msgspec.json.Decoder]`. Cache lifetime matches the decoder/client, not the process — no module-level state, no autouse cache-clear fixtures in tests.
+- **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` cached per-instance on `PydanticDecoder._adapters: dict[type, TypeAdapter]` (populated lazily on first `_get_adapter()` call); the msgspec adapter mirrors the pattern with `MsgspecDecoder._msgspec_decoders: dict[type, msgspec.json.Decoder]`. Cache lifetime matches the decoder/client, not the process — no module-level state, no autouse cache-clear fixtures in tests. Both built-ins' `_get_adapter`/`_get_msgspec_decoder` and `can_decode` share one memoizing helper (`decoders/_caching.py:_get_or_build`).
+- **Unhashable models:** a `model` that isn't hashable (e.g. `Annotated[X, unhashable_metadata]`) bypasses the cache entirely — every call builds fresh, uncached. `can_decode`'s verdict reflects true buildability regardless of hashability: an unhashable model that the underlying library can actually build a schema/decoder for returns `True`, not `False`. Hashability is orthogonal to whether a model can be decoded.

planning/changes/2026-07-13.05-decoders-shared-memoizing-cache.md

@@ -0,0 +1,289 @@
+---
+summary: Share the decoders' memoizing get-or-build cache pattern (4 duplicated sites across pydantic.py/msgspec.py) via one generic _get_or_build function, and drop decode()'s now-dead TypeError fallback.
+---
+
+# Design: Share the decoders' memoizing cache pattern
+
+## Summary
+
+`src/httpware/decoders/pydantic.py` and `decoders/msgspec.py` each
+implement the identical "check a `dict[type, V]` cache, build-and-memoize
+on miss, unhashable models bypass the cache entirely" pattern at two sites
+per file: the object cache (`_get_adapter`/`_get_msgspec_decoder`) and the
+verdict cache (`can_decode`). This change extracts one generic
+`_get_or_build` function into a new `decoders/_caching.py`, used at all
+four sites, and simplifies `decode()` in both classes by dropping a
+try/except that becomes dead code once the cache handles the unhashable
+case internally. Two existing tests that mocked the wrong layer are
+corrected to test the real mechanism. **One intentional, incidental
+behavior fix**: `can_decode` now agrees with `decode` for genuinely
+unhashable, schema-buildable models (see Motivation) — everything else is
+behavior-preserving.
+
+## Motivation
+
+- **Four duplicated cache sites, not two.** `PydanticDecoder._get_adapter`
+  (`:45-50`) and `MsgspecDecoder._get_msgspec_decoder` (`:67-72`) are
+  structurally identical (differing only in what gets built and which
+  dict holds it). `PydanticDecoder.can_decode` (`:52-67`) and
+  `MsgspecDecoder.can_decode` (`:74-89`) are byte-identical (modulo
+  docstring wording) — same `try: cache.get(model) / except TypeError:
+  probe fresh`, same `if cached is not None: return cached`, same
+  probe-store-return.
+- **The sentinel trick is shared too.** Both cache flavors — `dict[type,
+  bool]` (verdicts) and `dict[type, TypeAdapter | Decoder]` (built
+  objects) — rely on the same fact: a real stored value (`True`/`False`,
+  or a built object) is never `None`, so `dict.get(model)` returning
+  `None` unambiguously means "not cached." One generic function serves
+  both flavors.
+- **Corroborating signal:** `tests/test_decoders_pydantic.py` and
+  `tests/test_decoders_msgspec.py` (241 lines each) independently
+  re-verify the same caching contract via `# noqa: SLF001` pokes at
+  `_adapters`/`_can_decode_results` and `_msgspec_decoders`/
+  `_can_decode_results` — the memoization concern isn't exercisable
+  through the public `ResponseDecoder` protocol alone, so tests reach
+  through the seam rather than past it.
+- **Deletion test:** delete `_get_or_build` and the identical
+  get-or-build logic reappears at all four call sites — real, load-bearing
+  logic, not a pass-through.
+- **A second, independent finding surfaced while re-reading both files
+  closely:** `decode()` in both classes has its own try/except `TypeError`
+  fallback (falls back to a fresh, uncached build) that exists solely to
+  handle the same unhashable-model case `_get_or_build` will now handle
+  internally. Traced through Seam B's contract
+  (`decoders/_resolver.py`'s `_DecoderResolver.resolve`): `decode()` only
+  ever runs on a model that already passed `can_decode`'s probe in the
+  same request, so `TypeAdapter(model)`/`msgspec.json.Decoder(model)`
+  construction cannot newly fail at `decode()` time — the only thing the
+  existing fallback ever guarded against was the cache-lookup
+  `TypeError`. Once that's handled inside `_get_or_build`, the fallback in
+  `decode()` becomes genuinely unreachable and is removed.
+- **Discovered during task review, verified empirically: a pre-existing
+  inconsistency between `can_decode` and `decode` for genuinely
+  unhashable models, which this refactor incidentally fixes.** Today,
+  `_probe_can_decode` calls `self._get_adapter(model)` inside a broad
+  `except Exception: return False`. For a genuinely unhashable model
+  (e.g. `typing.Annotated[int, []]` — confirmed unhashable, and confirmed
+  buildable by both pydantic and msgspec), the *old* `_get_adapter`/
+  `_get_msgspec_decoder` propagate the cache-lookup `TypeError` uncaught,
+  which that `except Exception` catches — so `can_decode` returns `False`
+  for *any* unhashable model today, regardless of whether the underlying
+  library could actually build a schema for it. Meanwhile `decode()`
+  already works fine for the same model (it has its own separate
+  fallback) — meaning today, `MissingDecoderError` can incorrectly fire
+  for a model `decode()` would have handled correctly. Once
+  `_get_or_build` swallows that same `TypeError` internally (which it
+  must, to keep `decode()` working for unhashable models — see the point
+  above), `_probe_can_decode`'s `try` no longer sees an exception for
+  this case, so `can_decode` now returns `True` — consistent with
+  `decode`'s actual capability. This is accepted as an intentional
+  incidental fix (see Non-goals and Testing).
+
+## Non-goals
+
+- No *unintentional* behavior change from the caller's perspective. One
+  intentional exception, accepted as a bonus fix rather than reverted:
+  `can_decode` now returns `True` (not `False`) for a genuinely unhashable
+  model that the underlying library can actually build a schema for —
+  matching what `decode` already does for the same model, closing a
+  pre-existing inconsistency. Every other input/output pair for
+  `can_decode`/`decode` is unchanged, including unhashable models handled
+  via a fresh, uncached build (still true — that mechanism is unchanged,
+  only the *verdict* changes for the specific unhashable-but-buildable
+  case).
+- Not extracting the tiny `if not is_X_installed: raise
+  ImportError(MESSAGE)` check duplicated in both `__init__`s (2 lines,
+  different flag and message per call — a shared helper would need both
+  passed as parameters, more ceremony than it replaces). Same judgment
+  already applied to `_internal/status.py`'s 3-line predicate functions
+  in the original architecture review.
+- Not touching `_probe_can_decode` (genuinely different per decoder — the
+  actual varying logic this refactor is careful to leave alone) or
+  `_contains_custom_type` (msgspec-specific, unrelated concern).
+- Not touching the `ResponseDecoder` protocol, `_DecoderResolver`, or
+  `_BoundDecoder` (Seam B's orchestration, already deep, unaffected).
+
+## Design
+
+### 1. `decoders/_caching.py` — new module
+
+```python
+"""Shared memoizing get-or-build cache for the decoder adapters (Seam C).
+
+Imports neither pydantic nor msgspec — safe for both decoder modules to
+import without crossing Seam C's per-extra import isolation.
+"""
+
+import typing
+
+
+V = typing.TypeVar("V")
+
+
+def _get_or_build(cache: dict[type, V], model: type, build: typing.Callable[[], V]) -> V:
+    """Return cache[model], building and memoizing it via build() on miss.
+
+    Unhashable models bypass the cache entirely: dict.get(model) raises
+    TypeError, in which case this returns a fresh, uncached build() every
+    call rather than raising. A real cached value (bool verdict or a
+    built adapter/decoder) is never None, so a None result unambiguously
+    means "not cached yet" — callers must not store None as a value.
+    """
+    try:
+        cached = cache.get(model)
+    except TypeError:
+        return build()
+    if cached is not None:
+        return cached
+    result = build()
+    cache[model] = result
+    return result
+```
+
+Placed alongside `decoders/_resolver.py` — both are private, cross-file
+modules inside `decoders/` shared by the seam's adapters, not moved to
+`_internal/` (which is for logic shared across otherwise-unrelated parts
+of the whole package, not decoders-specific plumbing).
+
+### 2. `PydanticDecoder` — four call sites become two, `decode()` simplifies
+
+```python
+from httpware.decoders._caching import _get_or_build
+
+# ...
+
+def _get_adapter(self, model: type[T]) -> "TypeAdapter[T]":
+    return _get_or_build(self._adapters, model, lambda: TypeAdapter(model))
+
+def can_decode(self, model: type) -> bool:
+    """Return True iff pydantic can build a schema for `model`.
+
+    The verdict is memoized per `model` so a rejection (which costs a
+    `PydanticSchemaGenerationError` round-trip) is not re-probed on every
+    dispatch. Unhashable models skip the cache and probe fresh.
+    """
+    return _get_or_build(self._can_decode_results, model, lambda: self._probe_can_decode(model))
+
+def decode(self, content: bytes, model: type[T]) -> T:
+    """Validate `content` as JSON against `model` in a single parse pass."""
+    adapter = self._get_adapter(model)
+    return adapter.validate_json(content)
+```
+
+`_probe_can_decode` is untouched.
+
+### 3. `MsgspecDecoder` — the same shape
+
+```python
+from httpware.decoders._caching import _get_or_build
+
+# ...
+
+def _get_msgspec_decoder(self, model: type[T]) -> "msgspec.json.Decoder[T]":
+    return _get_or_build(self._msgspec_decoders, model, lambda: msgspec.json.Decoder(model))
+
+def can_decode(self, model: type) -> bool:
+    """Return True iff msgspec natively understands `model` end-to-end.
+
+    The verdict is memoized per `model`: the probe below (an uncached
+    `type_info` call plus a recursive tree walk) runs once per type, not
+    on every dispatch. Unhashable models skip the cache and probe fresh.
+    """
+    return _get_or_build(self._can_decode_results, model, lambda: self._probe_can_decode(model))
+
+def decode(self, content: bytes, model: type[T]) -> T:
+    """Validate `content` as JSON against `model` in a single parse pass."""
+    decoder = self._get_msgspec_decoder(model)
+    return decoder.decode(content)
+```
+
+`_probe_can_decode` and `_contains_custom_type` are untouched.
+
+### 4. Test corrections (necessary, not optional)
+
+Two existing tests mock the wrong layer and must change, or they break
+once `decode()`'s fallback is removed:
+
+- `tests/test_decoders_pydantic.py::test_unhashable_model_falls_back_to_uncached_adapter`
+  (currently `:134-150`) mocks `PydanticDecoder._get_adapter` itself with
+  `side_effect=TypeError(...)` to exercise `decode()`'s fallback. Once
+  `_get_adapter` delegates to `_get_or_build`, it can no longer raise
+  `TypeError` for this reason, so the mock describes a scenario the real
+  method can't produce. Rewrite to mock the cache dict instead
+  (`decoder._adapters = MagicMock(); decoder._adapters.get.side_effect =
+  TypeError(...)`), mirroring the pattern the existing
+  `test_pydantic_can_decode_unhashable_model_does_not_raise` (`:236-241`)
+  already uses for the verdict cache. Same assertions
+  (`decoder.decode(b"42", int) == 42`, then a genuinely malformed payload
+  still raises `pydantic.ValidationError`).
+- `tests/test_decoders_msgspec.py::test_unhashable_model_falls_back_to_uncached_decoder`
+  (currently `:140-154`) — identical fix, mocking `decoder._msgspec_decoders`
+  instead of `MsgspecDecoder._get_msgspec_decoder`.
+
+Once fixed this way, both tests directly exercise `_get_or_build`'s
+`TypeError`-bypass path for the object-cache flavor, complementing the
+existing `can_decode`-unhashable tests (verdict-cache flavor) — decent
+direct coverage of the new shared function without a dedicated test
+file.
+
+**Everything else in both test files is unaffected**, verified by reading
+every test in both files:
+- `test_can_decode_returns_false_when_decoder_build_raises` (msgspec
+  `:130-137`) mocks `_get_msgspec_decoder` to test `_probe_can_decode`'s
+  own broad `except Exception: return False` — untouched, since
+  `_probe_can_decode` still calls `self._get_msgspec_decoder(model)` as a
+  method regardless of its internal implementation.
+- The `patch.object(decoder, "_get_adapter", wraps=decoder._get_adapter)`
+  spy tests (pydantic `:219`, `:229`) wrap the real method rather than
+  replacing it — call-count assertions hold unchanged since the number of
+  calls to `_get_adapter`/`_probe_can_decode` doesn't change.
+- `test_cache_invariance_*` (pydantic `:87-131`) patch the module-level
+  `TypeAdapter` name directly — independent of internal call structure.
+- `test_pydantic_can_decode_uses_cache`/`test_msgspec_can_decode_uses_cache`
+  check dict contents (`len(...) == 1`, membership) — unaffected, the
+  dicts are populated the same way.
+
+### 5. Minor comment accuracy fix
+
+`test_cache_invariance_concurrent_first_calls_threadpool`'s comment
+(pydantic `:127-130`) says "the get→set sequence in `_get_adapter` is
+not [atomic]" — the get→set sequence now lives in `_get_or_build`. Update
+the comment to name the new location.
+
+## Testing
+
+- **Parity net:** every other existing test in both files stays green
+  unchanged (see the itemized check above).
+- **Corrected tests:** the two rewritten unhashable-fallback tests, as
+  specified in Design §4.
+- No new dedicated test file for `_get_or_build` — the corrected tests
+  plus the existing `can_decode`-unhashable tests already exercise both
+  value-type flavors (bool, built object) of its `TypeError`-bypass path,
+  and every cache-hit/cache-miss branch is already exercised by the
+  existing cache-content and call-count assertions.
+- **New regression tests for the intentional behavior fix** (added after
+  task review surfaced the gap empirically): one test per decoder, using
+  a genuinely unhashable model (`typing.Annotated[int, []]` — confirmed
+  unhashable via `hash()`, confirmed buildable by both pydantic and
+  msgspec), asserting `can_decode(model) is True` and that `decode(...)`
+  on the same model succeeds and agrees — proving the two are consistent
+  post-fix. No existing test used a genuinely unhashable model through
+  the real (non-mocked) code path; the two corrected tests from Design §4
+  mock the cache dict with an otherwise-hashable model, which does not
+  exercise this interaction.
+- `just lint && just test` both clean; 100% coverage maintained.
+
+## Risk
+
+- **Silently breaking the two rewritten tests' original intent**
+  (unlikely × medium): the fix must preserve the same observable
+  contract (unhashable model still decodes correctly via a fresh,
+  uncached build) even though the mock target moves. *Mitigation:* keep
+  the same assertions, only change what's mocked and how.
+- **Generic typing friction** (`_get_or_build`'s `V` vs. `T` narrowing at
+  each call site) (unlikely × low): `self._adapters`/`self._msgspec_decoders`
+  are already typed `dict[type, X[typing.Any]]` today, so the existing
+  code already widens through `Any` in `_get_adapter`'s own return
+  statement before this change — no new widening is introduced.
+  *Mitigation:* `ty check` in the verification gate catches any real
+  regression.

src/httpware/decoders/_caching.py

@@ -0,0 +1,30 @@
+"""Shared memoizing get-or-build cache for the decoder adapters (Seam C).
+
+Imports neither pydantic nor msgspec — safe for both decoder modules to
+import without crossing Seam C's per-extra import isolation.
+"""
+
+import typing
+
+
+V = typing.TypeVar("V")
+
+
+def _get_or_build(cache: dict[type, V], model: type, build: typing.Callable[[], V]) -> V:
+    """Return cache[model], building and memoizing it via build() on miss.
+
+    Unhashable models bypass the cache entirely: dict.get(model) raises
+    TypeError, in which case this returns a fresh, uncached build() every
+    call rather than raising. A real cached value (bool verdict or a
+    built adapter/decoder) is never None, so a None result unambiguously
+    means "not cached yet" — callers must not store None as a value.
+    """
+    try:
+        cached = cache.get(model)
+    except TypeError:
+        return build()
+    if cached is not None:
+        return cached
+    result = build()
+    cache[model] = result
+    return result