typevet 0.1.0.dev1__tar.gz → 0.3.0__tar.gz

{typevet-0.1.0.dev1 → typevet-0.3.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: typevet
-Version: 0.1.0.dev1
-Summary: Type-safe structured generation under hexagonal architecture (llama.cpp first)
+Version: 0.3.0
+Summary: Typed answers from model next-token probabilities, on llama.cpp and vLLM
 Keywords: evaluation,gemma,json-schema,llama-cpp,llm,structured-generation,structured-output,type-safe,vllm
 Author: Alberto-Codes
 Author-email: Alberto-Codes <alberto.codes.dev@gmail.com>
@@ -20,6 +20,7 @@ Requires-Dist: httpx>=0.28
 Requires-Dist: jsonschema>=4.26.0
 Requires-Dist: structlog>=24.1
 Requires-Dist: typer>=0.12 ; extra == 'cli'
+Requires-Dist: judgevet>=0.14,<0.16 ; extra == 'judgevet'
 Requires-Python: >=3.12
 Project-URL: Homepage, https://github.com/Alberto-Codes/typevet
 Project-URL: Documentation, https://alberto-codes.github.io/typevet/
@@ -27,10 +28,12 @@ Project-URL: Repository, https://github.com/Alberto-Codes/typevet
 Project-URL: Issues, https://github.com/Alberto-Codes/typevet/issues
 Project-URL: Changelog, https://github.com/Alberto-Codes/typevet/blob/main/CHANGELOG.md
 Provides-Extra: cli
+Provides-Extra: judgevet
 Description-Content-Type: text/markdown
 
 [![CI](https://img.shields.io/github/actions/workflow/status/Alberto-Codes/typevet/ci.yml?branch=main&label=CI)](https://github.com/Alberto-Codes/typevet/actions/workflows/ci.yml)
 [![Docs](https://img.shields.io/github/actions/workflow/status/Alberto-Codes/typevet/docs.yml?branch=main&label=docs)](https://alberto-codes.github.io/typevet/)
+[![PyPI](https://img.shields.io/pypi/v/typevet)](https://pypi.org/project/typevet/)
 [![Python](https://img.shields.io/badge/python-3.12-blue)](https://github.com/Alberto-Codes/typevet/blob/main/pyproject.toml)
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![docs vetted](https://img.shields.io/badge/docs%20vetted-docvet-purple)](https://github.com/Alberto-Codes/docvet)
@@ -49,18 +52,17 @@ Read the documentation at <https://alberto-codes.github.io/typevet/>.
 
 ## Status
 
-- typevet is pre-1.0. The package version is `0.1.0`.
-- typevet is not on PyPI yet.
-  Build a wheel from a checkout and install it: see
-  [Install typevet](https://github.com/Alberto-Codes/typevet/blob/main/docs/how-to/install.md).
+- typevet is pre-1.0. The package version is `0.2.0`.
+- Install typevet from PyPI: `pip install typevet` or `uv add typevet`.
+  See [Install typevet](https://alberto-codes.github.io/typevet/how-to/install/).
 - typevet requires Python 3.12 or later.
-- Each backend has one tested model pin.
-  The receipts give the full pin and its limits.
+- Each tested pin has one receipt.
+  The receipt gives the full pin and its limits.
 
 | Backend | Tested pin | Receipt |
 |---|---|---|
 | vLLM | `vllm/vllm-openai:v0.30.0`, BF16 `google/gemma-4-31B-it`, one H100 80 GB | [#170](https://github.com/Alberto-Codes/typevet/issues/170#issuecomment-5884707915) |
-| llama.cpp | Build `b11223-4da633776`, local alias `gemma-4-31b-kv9-q4km-mm` | [#203](https://github.com/Alberto-Codes/typevet/issues/203#issuecomment-5882379255) |
+| llama.cpp | Build `b11223-4da633776`, local alias `gemma-4-31b-kv9-q4km-mm`: file `gemma-4-31b-24gib-kv9-decoder.gguf`, ftype `Q2_K - Medium` (not Q4_K_M), 16.0 GB, `--chat-template-file` override ([#233](https://github.com/Alberto-Codes/typevet/issues/233)) | [#203](https://github.com/Alberto-Codes/typevet/issues/203#issuecomment-5882379255) |
 | llama.cpp grammar | Build `b11243-fc07d781e`, Gemma 4 31B QAT Q4_0 GGUF | [#129](https://github.com/Alberto-Codes/typevet/issues/129#issuecomment-5892208050) |
 
 Performance: on one H100 at concurrency level 64, 480 Banking77 records took 12.1 s at 39.6 records/s.
@@ -75,8 +77,8 @@ The receipts are small samples.
 Get one offline typed judgment from a scripted fake. This step needs no model.
 
 ```bash
-uv sync
-uv run python -c "
+pip install typevet
+python -c "
 from typevet.domain import Noul
 from typevet.runtime import ScoringJudgmentAdapter
 from typevet.testing import ScriptedScoringFake

{typevet-0.1.0.dev1 → typevet-0.3.0}/README.md

@@ -1,5 +1,6 @@
 [![CI](https://img.shields.io/github/actions/workflow/status/Alberto-Codes/typevet/ci.yml?branch=main&label=CI)](https://github.com/Alberto-Codes/typevet/actions/workflows/ci.yml)
 [![Docs](https://img.shields.io/github/actions/workflow/status/Alberto-Codes/typevet/docs.yml?branch=main&label=docs)](https://alberto-codes.github.io/typevet/)
+[![PyPI](https://img.shields.io/pypi/v/typevet)](https://pypi.org/project/typevet/)
 [![Python](https://img.shields.io/badge/python-3.12-blue)](https://github.com/Alberto-Codes/typevet/blob/main/pyproject.toml)
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![docs vetted](https://img.shields.io/badge/docs%20vetted-docvet-purple)](https://github.com/Alberto-Codes/docvet)
@@ -18,18 +19,17 @@ Read the documentation at <https://alberto-codes.github.io/typevet/>.
 
 ## Status
 
-- typevet is pre-1.0. The package version is `0.1.0`.
-- typevet is not on PyPI yet.
-  Build a wheel from a checkout and install it: see
-  [Install typevet](https://github.com/Alberto-Codes/typevet/blob/main/docs/how-to/install.md).
+- typevet is pre-1.0. The package version is `0.2.0`.
+- Install typevet from PyPI: `pip install typevet` or `uv add typevet`.
+  See [Install typevet](https://alberto-codes.github.io/typevet/how-to/install/).
 - typevet requires Python 3.12 or later.
-- Each backend has one tested model pin.
-  The receipts give the full pin and its limits.
+- Each tested pin has one receipt.
+  The receipt gives the full pin and its limits.
 
 | Backend | Tested pin | Receipt |
 |---|---|---|
 | vLLM | `vllm/vllm-openai:v0.30.0`, BF16 `google/gemma-4-31B-it`, one H100 80 GB | [#170](https://github.com/Alberto-Codes/typevet/issues/170#issuecomment-5884707915) |
-| llama.cpp | Build `b11223-4da633776`, local alias `gemma-4-31b-kv9-q4km-mm` | [#203](https://github.com/Alberto-Codes/typevet/issues/203#issuecomment-5882379255) |
+| llama.cpp | Build `b11223-4da633776`, local alias `gemma-4-31b-kv9-q4km-mm`: file `gemma-4-31b-24gib-kv9-decoder.gguf`, ftype `Q2_K - Medium` (not Q4_K_M), 16.0 GB, `--chat-template-file` override ([#233](https://github.com/Alberto-Codes/typevet/issues/233)) | [#203](https://github.com/Alberto-Codes/typevet/issues/203#issuecomment-5882379255) |
 | llama.cpp grammar | Build `b11243-fc07d781e`, Gemma 4 31B QAT Q4_0 GGUF | [#129](https://github.com/Alberto-Codes/typevet/issues/129#issuecomment-5892208050) |
 
 Performance: on one H100 at concurrency level 64, 480 Banking77 records took 12.1 s at 39.6 records/s.
@@ -44,8 +44,8 @@ The receipts are small samples.
 Get one offline typed judgment from a scripted fake. This step needs no model.
 
 ```bash
-uv sync
-uv run python -c "
+pip install typevet
+python -c "
 from typevet.domain import Noul
 from typevet.runtime import ScoringJudgmentAdapter
 from typevet.testing import ScriptedScoringFake

{typevet-0.1.0.dev1 → typevet-0.3.0}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "typevet"
-version = "0.1.0.dev1"
-description = "Type-safe structured generation under hexagonal architecture (llama.cpp first)"
+version = "0.3.0"
+description = "Typed answers from model next-token probabilities, on llama.cpp and vLLM"
 readme = "README.md"
 license = "MIT"
 license-files = ["LICENSE"]
@@ -40,6 +40,7 @@ email = "alberto.codes.dev@gmail.com"
 
 [project.optional-dependencies]
 cli = ["typer>=0.12"]
+judgevet = ["judgevet>=0.14,<0.16"]
 
 [project.urls]
 Homepage = "https://github.com/Alberto-Codes/typevet"
@@ -119,7 +120,7 @@ convention = "google"
 ]
 "evals/src/typevet_evals/wheel_isolated.py" = ["S603"]
 "scripts/check_commit_msg.py" = ["S603"]
-"src/typevet/evaluation/datasets/partner_guard.py" = [
+"evals/src/typevet_evals/datasets/partner_guard.py" = [
     "S603",
     "S607",
 ]
@@ -149,7 +150,6 @@ addopts = [
     "-ra",
     "--strict-markers",
     "--strict-config",
-    "--showlocals",
     "-m",
     "not live",
 ]
@@ -186,7 +186,7 @@ name = "Hexagonal layers"
 type = "layers"
 layers = [
     "typevet.runtime",
-    "typevet.evaluation : typevet.adapters.inbound",
+    "typevet.adapters.inbound",
     "typevet.adapters.outbound",
     "typevet.adapters.diagnostics",
     "typevet.testing",
@@ -201,10 +201,56 @@ source_modules = ["typevet"]
 forbidden_modules = ["typevet_evals"]
 
 [[tool.importlinter.contracts]]
-name = "runtime_must_not_import_evaluation"
+name = "Only the judgevet bridge imports judgevet"
 type = "forbidden"
-source_modules = ["typevet.runtime"]
-forbidden_modules = ["typevet.evaluation"]
+source_modules = [
+    "typevet",
+    "typevet_evals",
+]
+forbidden_modules = ["judgevet"]
+ignore_imports = ["typevet.adapters.inbound.judgevet -> judgevet"]
+
+[[tool.importlinter.contracts]]
+name = "No typevet module imports the judgevet bridge"
+type = "forbidden"
+source_modules = [
+    "typevet",
+    "typevet.**",
+    "typevet_evals",
+    "typevet_evals.**",
+]
+forbidden_modules = ["typevet.adapters.inbound.judgevet"]
+as_packages = false
+
+[[tool.importlinter.contracts]]
+name = "Evaluation families"
+type = "layers"
+layers = [
+    "typevet_evals.cli | typevet_evals.gemma_native_vision_wheel_smoke | typevet_evals.psai_vision_probability_evidence | typevet_evals.wheel_isolated",
+    "typevet_evals.throughput",
+    "typevet_evals.vllm_acceptance",
+    "typevet_evals.instruction_variant | typevet_evals.cord",
+    "typevet_evals.psai_vision_consumer | typevet_evals.outcome_replay_metrics",
+    "typevet_evals.runner | typevet_evals.tpjep | typevet_evals.experiment_identity | typevet_evals.face_match",
+    "typevet_evals.datasets",
+]
+
+[[tool.importlinter.contracts]]
+name = "Model framing stays off the serving backends"
+type = "forbidden"
+source_modules = ["typevet.adapters.outbound.gemma"]
+forbidden_modules = [
+    "typevet.adapters.outbound.llama_cpp",
+    "typevet.adapters.outbound.vllm",
+]
+
+[[tool.importlinter.contracts]]
+name = "Serving backends stay independent"
+type = "independence"
+modules = [
+    "typevet.adapters.outbound.llama_cpp",
+    "typevet.adapters.outbound.vllm",
+]
 
 [[tool.importlinter.contracts]]
 name = "Fakes stay off the adapters"
@@ -257,6 +303,7 @@ coauthor-grandfather-through = "1b1d9a3cfde019c5dbfe164c1660f1d3be28b5ba"
 dev = [
     "docvet[griffe]>=1.16.0",
     "import-linter>=2.15",
+    "judgevet[mcp,conformance]>=0.15,<0.16",
     "pre-commit>=4.6.2",
     "pytest>=9.1.1",
     "pytest-cov>=7.1.0",

{typevet-0.1.0.dev1 → typevet-0.3.0}/pyproject.toml.orig

@@ -1,7 +1,7 @@
 [project]
 name = "typevet"
-version = "0.1.0.dev1"
-description = "Type-safe structured generation under hexagonal architecture (llama.cpp first)"
+version = "0.3.0"
+description = "Typed answers from model next-token probabilities, on llama.cpp and vLLM"
 readme = "README.md"
 license = "MIT"
 license-files = ["LICENSE"]
@@ -37,6 +37,7 @@ dependencies = [
 
 [project.optional-dependencies]
 cli = ["typer>=0.12"]
+judgevet = ["judgevet>=0.14,<0.16"]
 
 [project.urls]
 Homepage = "https://github.com/Alberto-Codes/typevet"
@@ -56,6 +57,7 @@ include = ["src/typevet/py.typed"]
 dev = [
     "docvet[griffe]>=1.16.0",
     "import-linter>=2.15",
+    "judgevet[mcp,conformance]>=0.15,<0.16",
     "pre-commit>=4.6.2",
     "pytest>=9.1.1",
     "pytest-cov>=7.1.0",
@@ -131,7 +133,7 @@ convention = "google"
 ]
 "evals/src/typevet_evals/wheel_isolated.py" = ["S603"]
 "scripts/check_commit_msg.py" = ["S603"]
-"src/typevet/evaluation/datasets/partner_guard.py" = ["S603", "S607"]
+"evals/src/typevet_evals/datasets/partner_guard.py" = ["S603", "S607"]
 
 [tool.ruff.lint.isort]
 known-first-party = ["typevet", "typevet_evals"]
@@ -152,7 +154,6 @@ addopts = [
     "-ra",
     "--strict-markers",
     "--strict-config",
-    "--showlocals",
     "-m",
     "not live",
 ]
@@ -177,9 +178,7 @@ name = "Hexagonal layers"
 type = "layers"
 layers = [
     "typevet.runtime",
-    # Evaluation sits beside the inbound adapters: the eval CLI is an inbound
-    # adapter over evaluation, and the live gate reads inbound settings.
-    "typevet.evaluation : typevet.adapters.inbound",
+    "typevet.adapters.inbound",
     "typevet.adapters.outbound",
     "typevet.adapters.diagnostics",
     "typevet.testing",
@@ -194,10 +193,51 @@ source_modules = ["typevet"]
 forbidden_modules = ["typevet_evals"]
 
 [[tool.importlinter.contracts]]
-name = "runtime_must_not_import_evaluation"
+# The judgevet bridge (#284) is the one module that may import judgevet.
+name = "Only the judgevet bridge imports judgevet"
 type = "forbidden"
-source_modules = ["typevet.runtime"]
-forbidden_modules = ["typevet.evaluation"]
+source_modules = ["typevet", "typevet_evals"]
+forbidden_modules = ["judgevet"]
+ignore_imports = ["typevet.adapters.inbound.judgevet -> judgevet"]
+
+[[tool.importlinter.contracts]]
+# No package init or other module imports the bridge, so bare typevet stays judgevet-free.
+name = "No typevet module imports the judgevet bridge"
+type = "forbidden"
+# Module by module: as packages, typevet.adapters.inbound would contain the bridge and be skipped.
+source_modules = ["typevet", "typevet.**", "typevet_evals", "typevet_evals.**"]
+forbidden_modules = ["typevet.adapters.inbound.judgevet"]
+as_packages = false
+
+[[tool.importlinter.contracts]]
+name = "Evaluation families"
+type = "layers"
+layers = [
+    "typevet_evals.cli | typevet_evals.gemma_native_vision_wheel_smoke | typevet_evals.psai_vision_probability_evidence | typevet_evals.wheel_isolated",
+    "typevet_evals.throughput",
+    "typevet_evals.vllm_acceptance",
+    "typevet_evals.instruction_variant | typevet_evals.cord",
+    "typevet_evals.psai_vision_consumer | typevet_evals.outcome_replay_metrics",
+    "typevet_evals.runner | typevet_evals.tpjep | typevet_evals.experiment_identity | typevet_evals.face_match",
+    "typevet_evals.datasets",
+]
+
+[[tool.importlinter.contracts]]
+name = "Model framing stays off the serving backends"
+type = "forbidden"
+source_modules = ["typevet.adapters.outbound.gemma"]
+forbidden_modules = [
+    "typevet.adapters.outbound.llama_cpp",
+    "typevet.adapters.outbound.vllm",
+]
+
+[[tool.importlinter.contracts]]
+name = "Serving backends stay independent"
+type = "independence"
+modules = [
+    "typevet.adapters.outbound.llama_cpp",
+    "typevet.adapters.outbound.vllm",
+]
 
 [[tool.importlinter.contracts]]
 name = "Fakes stay off the adapters"

{typevet-0.1.0.dev1 → typevet-0.3.0}/src/typevet/adapters/inbound/backend_settings.py

@@ -15,14 +15,19 @@ The configured key never reaches the caller in clear text. ``VllmSettings``
 leaves it out of ``repr``. The adapters that ``generation_adapter`` and
 ``async_vllm_generation_adapter`` return catch each ``GenerationError`` from
 ``generate``, and the port that ``open_judgment`` yields catches each one from
-``judge``. When the error text, an attribute such as a parsed payload, or the
-cause chain holds the raw or JSON-escaped key, the error is raised again as
-the same type with the key replaced by ``***`` and with no cause or context.
-A server that echoes the ``Authorization`` header therefore cannot put the key
-into a ``BackendHttpError``. Successful results are not changed.
-The HTTP clients are built the same way with or without a key, so environment
-proxy settings apply in both cases. Error messages name variables, never
-values, and a parse failure raises with no cause or context.
+``judge``. When a key is configured, each such error is raised again as a copy
+of the same type with no cause or context. In the copy, the raw or JSON-escaped
+key in the text or an attribute, such as a parsed payload, is replaced by
+``***``. Strings and bytes inside dicts, lists, tuples, sets and frozensets are
+masked too, and masked bytes stay bytes. The copy is made even when the key
+text is absent, because an httpx error in the cause chain holds the request
+headers. A server that echoes the ``Authorization`` header therefore cannot put
+the key into a ``BackendHttpError``. Without a key, errors pass through
+unchanged. Successful results are not changed. The HTTP clients are built the
+same way with or without a key, so environment proxy settings apply in both
+cases. ``TYPEVET_VLLM__TIMEOUT`` must be finite and positive, so ``nan`` and
+``inf`` fail. Error messages name variables, never values, and a parse failure
+raises with no cause or context.
 
 Examples:
     ```python
@@ -41,9 +46,11 @@ Examples:
 
 See Also:
     - [typevet.adapters.inbound.settings][]: ``TYPEVET_LLAMA__*`` settings
-    - [typevet.adapters.outbound.vllm_generation][]: vLLM generation adapter
-    - [typevet.adapters.outbound.vllm_generation_async][]: Async vLLM adapter
-    - [typevet.adapters.outbound.vllm_judgment_factory][]: vLLM judgment factory
+    - [typevet.adapters.outbound.llama_cpp][]: llama.cpp adapters and the Gemma
+      native vision factory module
+    - [typevet.adapters.outbound.vllm.generation][]: vLLM generation adapter
+    - [typevet.adapters.outbound.vllm.generation_async][]: Async vLLM adapter
+    - [typevet.adapters.outbound.vllm.judgment_factory][]: vLLM judgment factory
     - [typevet.adapters.diagnostics.redaction][]: ``REDACTED`` (``***``) mask
     - docs/reference/configuration.md: Environment variable reference
 """
@@ -51,6 +58,7 @@ See Also:
 from __future__ import annotations
 
 import json
+import math
 import os
 from collections.abc import Iterator, Mapping
 from contextlib import contextmanager
@@ -61,14 +69,14 @@ import httpx
 
 from typevet.adapters.diagnostics.redaction import REDACTED
 from typevet.adapters.inbound.settings import llama_cpp_adapter, load_llama_settings
-from typevet.adapters.outbound.gemma_native_vision_factory import (
+from typevet.adapters.outbound.llama_cpp.gemma_native_vision_factory import (
     GemmaNativeVisionSession,
     open_gemma_native_vision_judgment,
 )
-from typevet.adapters.outbound.llama_cpp import LlamaCppGenerationAdapter
-from typevet.adapters.outbound.vllm_generation import VllmGenerationAdapter
-from typevet.adapters.outbound.vllm_generation_async import AsyncVllmGenerationAdapter
-from typevet.adapters.outbound.vllm_judgment_factory import (
+from typevet.adapters.outbound.llama_cpp.generation import LlamaCppGenerationAdapter
+from typevet.adapters.outbound.vllm.generation import VllmGenerationAdapter
+from typevet.adapters.outbound.vllm.generation_async import AsyncVllmGenerationAdapter
+from typevet.adapters.outbound.vllm.judgment_factory import (
     VllmJudgmentSession,
     open_vllm_judgment,
 )
@@ -117,63 +125,54 @@ class VllmSettings:
     user_agent: str | None = None
 
 
-def _masked(value: Any, needles: tuple[str, ...]) -> Any:
-    """Replace each key form in ``value`` with ``MASK``.
+_PLAIN_CONTAINERS: tuple[type, ...] = (list, tuple, set, frozenset)
+
 
-    Strings are masked in place of each needle. Dicts, lists and tuples are
-    copied as plain containers with each key and item masked, so a parsed payload that holds the
-    key loses it. Other values are returned unchanged.
+def _masked_text(value: str | bytes, needles: tuple[str, ...]) -> str | bytes:
+    """Replace each needle in ``value`` with ``MASK``, keeping its type.
+
+    Bytes stay bytes: each needle and ``MASK`` are ASCII-encoded first, which
+    is safe because the key must be ASCII.
 
     Args:
-        value: String, container or other attribute value.
+        value: String or bytes to mask.
         needles: Key forms to replace.
 
     Returns:
-        The masked value.
+        The masked string or bytes.
     """
     if isinstance(value, str):
         for needle in needles:
             value = value.replace(needle, MASK)
         return value
-    if isinstance(value, dict):
-        return {_masked(k, needles): _masked(v, needles) for k, v in value.items()}
-    if isinstance(value, list):
-        return [_masked(item, needles) for item in value]
-    if isinstance(value, tuple):
-        return tuple(_masked(item, needles) for item in value)
+    for needle in needles:
+        value = value.replace(needle.encode("ascii"), MASK.encode("ascii"))
     return value
 
 
-def _texts(value: Any) -> Iterator[str]:
-    """Yield every string inside ``value``, including dict keys.
+def _masked(value: Any, needles: tuple[str, ...]) -> Any:
+    """Replace each key form in ``value`` with ``MASK``.
+
+    Strings and bytes are masked by ``_masked_text`` and keep their type.
+    Dicts, lists, tuples, sets and frozensets are copied as plain containers
+    of the same kind with each key and item masked, so a parsed payload that
+    holds the key loses it. Other values are returned unchanged.
 
     Args:
-        value: String, container or other attribute value.
+        value: String, bytes, container or other attribute value.
+        needles: Key forms to replace.
 
-    Yields:
-        Each string that ``_masked`` would mask.
+    Returns:
+        The masked value.
     """
-    if isinstance(value, str):
-        yield value
-    elif isinstance(value, dict):
-        for key, item in value.items():
-            yield from _texts(key)
-            yield from _texts(item)
-    elif isinstance(value, list | tuple):
-        for item in value:
-            yield from _texts(item)
-
-
-def _chain_text(exc: BaseException) -> str:
-    parts: list[str] = []
-    seen: set[int] = set()
-    current: BaseException | None = exc
-    while current is not None and id(current) not in seen:
-        seen.add(id(current))
-        parts += [str(current), repr(current)]
-        parts += [text for v in vars(current).values() for text in _texts(v)]
-        current = current.__cause__ or current.__context__
-    return "\n".join(parts)
+    if isinstance(value, (str, bytes)):
+        return _masked_text(value, needles)
+    if isinstance(value, dict):
+        return {_masked(k, needles): _masked(v, needles) for k, v in value.items()}
+    for kind in _PLAIN_CONTAINERS:
+        if isinstance(value, kind):
+            return kind(_masked(item, needles) for item in value)
+    return value
 
 
 def _masked_error(exc: GenerationError, needles: tuple[str, ...]) -> GenerationError:
@@ -188,7 +187,8 @@ def _masked_error(exc: GenerationError, needles: tuple[str, ...]) -> GenerationE
 
     Returns:
         A new error whose arguments and attributes are masked, including
-        strings inside dict, list and tuple values such as a payload.
+        strings and bytes inside dict, list, tuple, set and frozenset
+        values such as a payload.
     """
     masked = type(exc).__new__(type(exc))
     masked.args = _masked(exc.args, needles)
@@ -206,17 +206,20 @@ def _key_needles(key: str | None) -> tuple[str, ...]:
 def _masked_if_keyed(
     exc: GenerationError, needles: tuple[str, ...]
 ) -> GenerationError | None:
-    """Return a masked copy of ``exc`` when its chain holds a key form.
+    """Return a masked copy of ``exc`` when a key is configured.
+
+    The copy is made whether or not the key text appears in ``exc``. The
+    cause chain can hold the key where no text check sees it, for example in
+    the headers of an httpx request, so the copy drops the chain in all cases.
 
     Args:
         exc: Error raised by a vLLM adapter.
-        needles: Key forms from ``_key_needles``.
+        needles: Key forms from ``_key_needles``, or empty without a key.
 
     Returns:
-        The masked copy, or ``None`` when no key form is present.
+        The masked copy, or ``None`` when no key is configured.
     """
-    chain = _chain_text(exc)
-    if not any(needle in chain for needle in needles):
+    if not needles:
         return None
     return _masked_error(exc, needles)
 
@@ -258,7 +261,7 @@ class _KeyMaskingJudgmentPort:
             The response from the wrapped port, unchanged.
 
         Raises:
-            GenerationError: The port error, masked when it holds the key.
+            GenerationError: The port error, as a masked copy when a key is set.
         """
         try:
             return self._port.judge(state, questions, model, media=media)
@@ -291,10 +294,11 @@ class _ClientOwningVllmAdapter(VllmGenerationAdapter):
     def generate(self, request: GenerationRequest) -> GenerationResult:
         """Generate, and mask the configured key in any raised error.
 
-        A successful result is returned unchanged. ``_masked_if_keyed``
-        checks each error: one whose text or cause chain holds the key is
-        raised again as the same type with the key replaced by ``MASK`` and
-        with no cause or context.
+        A successful result is returned unchanged. When a key is configured,
+        ``_masked_if_keyed`` copies each error as the same type with the key
+        replaced by ``MASK`` and with no cause or context. The copy is raised
+        outside the ``except`` block, so the original error is not reachable
+        from it.
 
         Args:
             request: Prompt, schema and served model name.
@@ -303,7 +307,7 @@ class _ClientOwningVllmAdapter(VllmGenerationAdapter):
             The result from ``VllmGenerationAdapter.generate``.
 
         Raises:
-            GenerationError: The adapter error, masked when it holds the key.
+            GenerationError: The adapter error, as a masked copy when a key is set.
         """
         try:
             return super().generate(request)
@@ -348,10 +352,10 @@ class _ClientOwningAsyncVllmAdapter(AsyncVllmGenerationAdapter):
     async def generate(self, request: GenerationRequest) -> GenerationResult:
         """Generate, and mask the configured key in any raised error.
 
-        Masking follows ``_ClientOwningVllmAdapter.generate``: an error whose
-        text, attributes or cause chain holds the key is raised again as the
-        same type with the key replaced by ``MASK`` and with no cause or
-        context. A successful result is returned unchanged.
+        Masking follows ``_ClientOwningVllmAdapter.generate``: when a key is
+        configured, each error is raised again as a copy of the same type
+        with the key replaced by ``MASK`` and with no cause or context. A
+        successful result is returned unchanged.
 
         Args:
             request: Prompt, schema and served model name.
@@ -360,7 +364,7 @@ class _ClientOwningAsyncVllmAdapter(AsyncVllmGenerationAdapter):
             The result from ``AsyncVllmGenerationAdapter.generate``.
 
         Raises:
-            GenerationError: The adapter error, masked when it holds the key.
+            GenerationError: The adapter error, as a masked copy when a key is set.
             RuntimeError: When the call runs on a different event loop from
                 the first call. It is not masked, because it holds no key.
         """
@@ -386,8 +390,8 @@ def _read_timeout(source: Mapping[str, str], name: str) -> float:
         timeout = float(raw)
     except ValueError:
         timeout = 0.0
-    if timeout <= 0:
-        msg = f"{name} must be a positive number of seconds"
+    if not math.isfinite(timeout) or timeout <= 0:
+        msg = f"{name} must be a finite positive number of seconds"
         raise ValueError(msg) from None
     return timeout
 
@@ -449,7 +453,8 @@ def load_vllm_settings(environ: Mapping[str, str] | None = None) -> VllmSettings
 
     Raises:
         ValueError: When ``TYPEVET_VLLM__BASE_URL`` or ``TYPEVET_VLLM__MODEL``
-            is missing, ``TYPEVET_VLLM__TIMEOUT`` is not a positive number,
+            is missing, ``TYPEVET_VLLM__TIMEOUT`` is not a finite positive
+            number,
             ``TYPEVET_VLLM__MAX_CONCURRENCY`` is not a positive integer, or
             ``TYPEVET_VLLM__API_KEY`` holds a non-ASCII character.
     """