fix(evidence): harden canonical corpus contracts

.superpowers/sdd/evidence-task-1-report.md

@@ -41,3 +41,33 @@ did not exist. After implementation, the focused suite passed.
   treat these value objects as immutable and construct replacements rather than mutate collections.
 - The adapter and normalization tasks should preserve the credential-free boundary by passing only
   these records beyond acquisition.
+
+## Review hardening follow-up
+
+All six binding review areas were addressed in a separate TDD pass:
+
+- JSON metadata is recursively converted to immutable `FrozenDict`/tuple values while retaining
+  stable object/array JSON serialization. Manifest document and chunk collections are tuples.
+- Secret-key matching now normalizes camelCase and punctuation. It rejects credential-specific
+  names (passwords, API keys, access/refresh tokens, client/private keys, session cookies and
+  authorization) recursively, while deliberate benign labels such as generic `token` and `secret`
+  remain valid.
+- Canonical URIs require a scheme and reject userinfo or credential-bearing query parameters.
+- Namespaced IDs, SHA-256 content hashes, timezone-aware UTC timestamps, embedding/vector
+  compatibility, unique IDs, chunk referential/provenance integrity, contiguous per-document
+  ordinals and pipeline-version consistency are validated. Nested Pydantic instances are always
+  revalidated so `model_copy(update=...)` cannot bypass a manifest boundary.
+- Acquired arbitrary bytes have explicit base64 JSON encoding and validation, covered by a JSON
+  round-trip test.
+- `EvidenceSourceError` classifies transient/retryable versus permanent failures and exposes only
+  recursively immutable, credential-screened JSON details.
+
+Follow-up verification:
+
+- Focused contract suite: 39 passed.
+- Focused Ruff: passed.
+- Harness excluding Docker-backed L0 and the network-dependent wheel packaging test: 470 passed,
+  5 deselected.
+- Fresh unrestricted harness attempt: 479 passed, 5 deselected; the same environmental boundary
+  remains (47 Docker socket setup errors, four Docker parity failures, one isolated `uv build`
+  network failure).

harness/tests/test_corpus_models.py

@@ -1,4 +1,4 @@
-from datetime import UTC, datetime
+from datetime import UTC, datetime, timedelta, timezone
 
 import pytest
 from pydantic import ValidationError
@@ -12,11 +12,11 @@ def document(source_uri: str = "https://host/a.md") -> CanonicalDocument:
         source_id="source:a",
         source_uri=source_uri,
         source_fingerprint="etag:abc",
-        content_hash="sha256:def",
+        content_hash=f"sha256:{'d' * 64}",
         title="A",
         content="# A",
         media_type="text/markdown",
-        pipeline_version="normalize-v1",
+        pipeline_version="evidence-v1",
     )
 
 
@@ -26,9 +26,9 @@ def chunk() -> CanonicalChunk:
         document_id="doc:abc",
         ordinal=0,
         content="# A",
-        content_hash="sha256:def",
+        content_hash=f"sha256:{'e' * 64}",
         source_uri="https://host/a.md",
-        pipeline_version="chunk-v1",
+        pipeline_version="evidence-v1",
     )
 
 
@@ -65,17 +65,19 @@ def test_canonical_records_are_frozen(model):
         model.pipeline_version = "changed"  # type: ignore[misc]
 
 
-def test_manifest_collections_have_independent_defaults():
+def test_manifest_collections_are_immutable_tuples_with_json_arrays():
     first = CorpusManifest(
-        manifest_id="one", created_at=datetime.now(UTC), pipeline_version="v1"
+        manifest_id="manifest:one", created_at=datetime.now(UTC), pipeline_version="v1"
     )
     second = CorpusManifest(
-        manifest_id="two", created_at=datetime.now(UTC), pipeline_version="v1"
+        manifest_id="manifest:two", created_at=datetime.now(UTC), pipeline_version="v1"
     )
 
-    first.documents.append(document())
-
-    assert second.documents == []
+    with pytest.raises(AttributeError):
+        first.documents.append(document())
+    assert first.documents == ()
+    assert second.documents == ()
+    assert '"documents":[]' in first.model_dump_json()
 
 
 def test_manifest_validates_embedding_compatibility_fields():
@@ -104,3 +106,69 @@ def test_canonical_metadata_rejects_secrets_and_non_json_values():
             pipeline_version="v1",
             metadata={"bad": object()},
         )
+
+
+def test_manifest_rejects_duplicate_ids_and_source_ids():
+    first = document()
+    duplicate_source = first.model_copy(
+        update={"document_id": "doc:other", "source_uri": "https://host/b.md"}
+    )
+    with pytest.raises(ValidationError, match="source_id"):
+        CorpusManifest(pipeline_version="evidence-v1", documents=[first, duplicate_source])
+
+    with pytest.raises(ValidationError, match="chunk_id"):
+        CorpusManifest(
+            pipeline_version="evidence-v1", documents=[first], chunks=[chunk(), chunk()]
+        )
+
+
+def test_manifest_rejects_orphan_noncontiguous_and_inconsistent_chunks():
+    with pytest.raises(ValidationError, match="unknown document"):
+        CorpusManifest(pipeline_version="evidence-v1", chunks=[chunk()])
+
+    second = chunk().model_copy(update={"chunk_id": "chunk:abc:2", "ordinal": 2})
+    with pytest.raises(ValidationError, match="contiguous"):
+        CorpusManifest(
+            pipeline_version="evidence-v1", documents=[document()], chunks=[chunk(), second]
+        )
+
+    wrong_uri = chunk().model_copy(update={"source_uri": "https://host/wrong.md"})
+    with pytest.raises(ValidationError, match="source_uri"):
+        CorpusManifest(
+            pipeline_version="evidence-v1", documents=[document()], chunks=[wrong_uri]
+        )
+
+
+def test_manifest_rejects_inconsistent_pipeline_versions():
+    wrong = document().model_copy(update={"pipeline_version": "other-v1"})
+    with pytest.raises(ValidationError, match="pipeline_version"):
+        CorpusManifest(pipeline_version="evidence-v1", documents=[wrong])
+
+
+def test_vector_generation_requires_embedding_compatibility():
+    with pytest.raises(ValidationError, match="vector_generation"):
+        CorpusManifest(pipeline_version="evidence-v1", vector_generation="generation:one")
+
+
+@pytest.mark.parametrize(
+    ("field", "value"),
+    [
+        ("document_id", "not-namespaced"),
+        ("content_hash", "sha256:not-hex"),
+        ("source_uri", "https://user:pass@host/a"),
+        ("source_uri", "https://host/a?refresh_token=secret"),
+    ],
+)
+def test_canonical_document_rejects_malformed_or_sensitive_provenance(field, value):
+    with pytest.raises(ValidationError):
+        CanonicalDocument.model_validate({**document().model_dump(), field: value})
+
+
+def test_manifest_datetimes_are_aware_and_normalized_to_utc():
+    with pytest.raises(ValidationError, match="timezone-aware"):
+        CorpusManifest(created_at=datetime(2026, 7, 12), pipeline_version="evidence-v1")
+
+    plus_two = datetime(2026, 7, 12, 12, tzinfo=timezone(timedelta(hours=2)))
+    manifest = CorpusManifest(created_at=plus_two, pipeline_version="evidence-v1")
+    assert manifest.created_at.tzinfo is UTC
+    assert manifest.created_at.hour == 10

harness/tests/test_evidence_port_contract.py

@@ -1,9 +1,15 @@
-from datetime import UTC, datetime
+from datetime import UTC, datetime, timedelta, timezone
 
 import pytest
 from pydantic import ValidationError
 
-from tht.ports.evidence import AcquiredDocument, EvidenceSource, SourceObject
+from tht.ports.evidence import (
+    AcquiredDocument,
+    EvidenceSource,
+    EvidenceSourceError,
+    EvidenceSourceErrorCategory,
+    SourceObject,
+)
 
 
 class StubSource:
@@ -11,7 +17,7 @@ class StubSource:
         return iter(
             [
                 SourceObject(
-                    source_id="handbook",
+                    source_id="source:handbook",
                     uri="https://host/handbook.md",
                     fingerprint="sha256:abc",
                 )
@@ -35,35 +41,164 @@ def test_runtime_checkable_source_protocol():
 
 
 def test_source_objects_are_frozen_and_metadata_defaults_are_independent():
-    first = SourceObject(source_id="a", uri="file:///a", fingerprint="sha256:a")
-    second = SourceObject(source_id="b", uri="file:///b", fingerprint="sha256:b")
+    first = SourceObject(source_id="source:a", uri="file:///a", fingerprint="sha256:a")
+    second = SourceObject(source_id="source:b", uri="file:///b", fingerprint="sha256:b")
 
     with pytest.raises(ValidationError):
         first.uri = "file:///changed"  # type: ignore[misc]
-    first.metadata["owner"] = "team-a"
+    with pytest.raises(TypeError):
+        first.metadata["owner"] = "team-a"
     assert second.metadata == {}
 
 
-@pytest.mark.parametrize("metadata", [{"api_key": "secret"}, {"auth": {"token": "secret"}}])
-def test_source_metadata_rejects_credentials(metadata):
+@pytest.mark.parametrize(
+    "key",
+    [
+        "password",
+        "PassWd",
+        "api_key",
+        "x-api-key",
+        "accessToken",
+        "refresh.token",
+        "client secret",
+        "privateKey",
+        "session_cookie",
+        "Authorization",
+    ],
+)
+def test_source_metadata_rejects_credential_specific_keys(key):
     with pytest.raises(ValidationError, match="credential-like"):
         SourceObject(
-            source_id="a", uri="https://host/a", fingerprint="etag:abc", metadata=metadata
+            source_id="source:a",
+            uri="https://host/a",
+            fingerprint="etag:abc",
+            metadata={"nested": [{key: "secret"}]},
         )
 
 
+def test_source_metadata_allows_benign_generic_token_and_secret_labels():
+    source = SourceObject(
+        source_id="source:a",
+        uri="https://host/a",
+        fingerprint="etag:abc",
+        metadata={"token": "word count token", "secret": False},
+    )
+
+    assert source.metadata["token"] == "word count token"
+
+
+def test_nested_metadata_is_recursively_immutable_and_serializes_as_json():
+    source = SourceObject(
+        source_id="source:a",
+        uri="https://host/a",
+        fingerprint="etag:abc",
+        metadata={"nested": {"items": [1, {"ok": True}]}},
+    )
+
+    with pytest.raises(TypeError):
+        source.metadata["nested"]["items"][1]["ok"] = False
+    assert '"items":[1,{"ok":true}]' in source.model_dump_json()
+
+
+@pytest.mark.parametrize(
+    "uri",
+    [
+        "https://user:pass@host/a",
+        "https://host/a?api_key=secret",
+        "https://host/a?accessToken=secret",
+    ],
+)
+def test_source_uri_rejects_embedded_credentials(uri):
+    with pytest.raises(ValidationError, match="credentials"):
+        SourceObject(source_id="source:a", uri=uri, fingerprint="etag:abc")
+
+
 def test_source_metadata_must_be_json_safe():
     with pytest.raises(ValidationError):
         SourceObject(
-            source_id="a",
+            source_id="source:a",
             uri="file:///a",
             fingerprint="sha256:a",
             metadata={"path": object()},
         )
 
 
+def test_source_identity_and_fingerprint_must_be_namespaced():
+    with pytest.raises(ValidationError, match="namespaced"):
+        SourceObject(source_id="plain", uri="file:///a", fingerprint="sha256:a")
+    with pytest.raises(ValidationError, match="namespaced"):
+        SourceObject(source_id="source:a", uri="file:///a", fingerprint="plain")
+
+
 def test_acquired_document_does_not_accept_credentials_as_extra_fields():
-    item = SourceObject(source_id="a", uri="https://host/a", fingerprint="etag:abc")
+    item = SourceObject(source_id="source:a", uri="https://host/a", fingerprint="etag:abc")
 
     with pytest.raises(ValidationError):
         AcquiredDocument(source=item, content=b"a", api_key="secret")
+
+
+def test_acquired_binary_content_has_explicit_json_round_trip():
+    item = SourceObject(source_id="source:a", uri="https://host/a", fingerprint="etag:abc")
+    acquired = AcquiredDocument(source=item, content=b"\x00\xffbinary\x80")
+
+    payload = acquired.model_dump_json()
+    restored = AcquiredDocument.model_validate_json(payload)
+
+    assert restored.content == acquired.content
+    assert "binary" not in payload
+
+
+def test_datetimes_must_be_aware_and_are_normalized_to_utc():
+    with pytest.raises(ValidationError, match="timezone-aware"):
+        SourceObject(
+            source_id="source:a",
+            uri="https://host/a",
+            fingerprint="etag:abc",
+            modified_at=datetime(2026, 7, 12),
+        )
+
+    source = SourceObject(
+        source_id="source:a",
+        uri="https://host/a",
+        fingerprint="etag:abc",
+        modified_at=datetime(2026, 7, 12, 4, tzinfo=timezone(timedelta(hours=2))),
+    )
+    assert source.modified_at.tzinfo is UTC
+    assert source.modified_at.hour == 2
+
+    acquired = AcquiredDocument(
+        source=source,
+        content=b"a",
+        acquired_at=datetime(2026, 7, 12, 2, tzinfo=UTC) + timedelta(hours=0),
+    )
+    assert acquired.acquired_at.utcoffset() == timedelta(0)
+
+
+def test_source_errors_are_typed_retryable_and_safe():
+    transient = EvidenceSourceError(
+        "remote source unavailable",
+        category=EvidenceSourceErrorCategory.TRANSIENT,
+        details={"status": 503},
+    )
+    permanent = EvidenceSourceError(
+        "unsupported media type",
+        category=EvidenceSourceErrorCategory.PERMANENT,
+    )
+
+    assert transient.retryable is True
+    assert permanent.retryable is False
+    assert transient.details["status"] == 503
+    with pytest.raises(TypeError):
+        transient.details["status"] = 200
+    with pytest.raises(ValueError, match="credential-like"):
+        EvidenceSourceError(
+            "bad",
+            category=EvidenceSourceErrorCategory.PERMANENT,
+            details={"apiKey": "must-not-leak"},
+        )
+    with pytest.raises(ValidationError):
+        EvidenceSourceError(
+            "bad",
+            category=EvidenceSourceErrorCategory.PERMANENT,
+            details={"not_json": object()},
+        )

harness/tht/corpus/models.py

@@ -1,64 +1,135 @@
 """Immutable records emitted by the Evidence preprocessing pipeline."""
 
+import re
 from datetime import UTC, datetime
 
 from pydantic import BaseModel, ConfigDict, Field, JsonValue, field_validator, model_validator
 
-from tht.ports.evidence import _reject_credentials
+from tht.ports.evidence import (
+    normalize_aware_datetime,
+    validate_canonical_uri,
+    validate_namespaced_value,
+    validate_safe_metadata,
+)
+
+
+_NAMESPACED_ID = re.compile(r"^[a-z][a-z0-9_-]*:[A-Za-z0-9._:-]+$")
+_SHA256 = re.compile(r"^sha256:[0-9a-f]{64}$")
+
+
+def _validate_namespaced_id(value: str) -> str:
+    if not _NAMESPACED_ID.fullmatch(value):
+        raise ValueError("identifier must be namespaced as '<kind>:<stable-value>'")
+    return value
+
+
+def _validate_hash(value: str) -> str:
+    if not _SHA256.fullmatch(value):
+        raise ValueError("content hash must be 'sha256:' followed by 64 lowercase hex digits")
+    return value
 
 
 class _CanonicalValue(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
+    model_config = ConfigDict(
+        frozen=True, extra="forbid", validate_default=True, revalidate_instances="always"
+    )
 
 
 class _WithMetadata(_CanonicalValue):
     metadata: dict[str, JsonValue] = Field(default_factory=dict)
-
-    @field_validator("metadata")
-    @classmethod
-    def metadata_has_no_credentials(cls, value: dict[str, JsonValue]) -> dict[str, JsonValue]:
-        _reject_credentials(value)
-        return value
+    _frozen_metadata = field_validator("metadata")(validate_safe_metadata)
 
 
 class CanonicalDocument(_WithMetadata):
-    document_id: str = Field(min_length=1)
-    source_id: str = Field(min_length=1)
-    source_uri: str = Field(min_length=1)
+    document_id: str
+    source_id: str
+    source_uri: str
     source_fingerprint: str = Field(min_length=1)
-    content_hash: str = Field(min_length=1)
+    content_hash: str
     title: str = ""
     content: str
     media_type: str = "text/plain"
     modified_at: datetime | None = None
     pipeline_version: str = Field(min_length=1)
 
+    _document_id = field_validator("document_id")(_validate_namespaced_id)
+    _source_id = field_validator("source_id")(_validate_namespaced_id)
+    _source_uri = field_validator("source_uri")(validate_canonical_uri)
+    _source_fingerprint = field_validator("source_fingerprint")(validate_namespaced_value)
+    _content_hash = field_validator("content_hash")(_validate_hash)
+    _modified_at = field_validator("modified_at")(normalize_aware_datetime)
+
 
 class CanonicalChunk(_WithMetadata):
-    chunk_id: str = Field(min_length=1)
-    document_id: str = Field(min_length=1)
+    chunk_id: str
+    document_id: str
     ordinal: int = Field(ge=0)
     content: str
-    content_hash: str = Field(min_length=1)
-    source_uri: str = Field(min_length=1)
+    content_hash: str
+    source_uri: str
     pipeline_version: str = Field(min_length=1)
 
+    _chunk_id = field_validator("chunk_id")(_validate_namespaced_id)
+    _document_id = field_validator("document_id")(_validate_namespaced_id)
+    _content_hash = field_validator("content_hash")(_validate_hash)
+    _source_uri = field_validator("source_uri")(validate_canonical_uri)
+
 
 class CorpusManifest(_WithMetadata):
-    """Description of one publishable canonical/vector generation."""
+    """Description of one internally consistent publishable generation."""
 
     schema_version: int = Field(default=1, ge=1)
     manifest_id: str | None = None
     created_at: datetime = Field(default_factory=lambda: datetime.now(UTC))
-    pipeline_version: str = Field(default="1", min_length=1)
+    pipeline_version: str = Field(default="evidence-v1", min_length=1)
     embedding_model: str | None = None
     embedding_dimensions: int | None = Field(default=None, gt=0)
     vector_generation: str | None = None
-    documents: list[CanonicalDocument] = Field(default_factory=list)
-    chunks: list[CanonicalChunk] = Field(default_factory=list)
+    documents: tuple[CanonicalDocument, ...] = Field(default_factory=tuple)
+    chunks: tuple[CanonicalChunk, ...] = Field(default_factory=tuple)
+
+    _manifest_id = field_validator("manifest_id")(
+        lambda value: _validate_namespaced_id(value) if value is not None else None
+    )
+    _vector_generation = field_validator("vector_generation")(
+        lambda value: _validate_namespaced_id(value) if value is not None else None
+    )
+    _created_at = field_validator("created_at")(normalize_aware_datetime)
 
     @model_validator(mode="after")
-    def embedding_fields_are_complete(self) -> "CorpusManifest":
+    def validate_generation(self) -> "CorpusManifest":
         if (self.embedding_model is None) != (self.embedding_dimensions is None):
             raise ValueError("embedding_model and embedding_dimensions must be set together")
+        if self.vector_generation is not None and self.embedding_model is None:
+            raise ValueError("vector_generation requires embedding model and dimension compatibility")
+
+        document_ids = [document.document_id for document in self.documents]
+        source_ids = [document.source_id for document in self.documents]
+        chunk_ids = [chunk.chunk_id for chunk in self.chunks]
+        self._require_unique("document_id", document_ids)
+        self._require_unique("source_id", source_ids)
+        self._require_unique("chunk_id", chunk_ids)
+
+        documents = {document.document_id: document for document in self.documents}
+        ordinals: dict[str, list[int]] = {}
+        for document in self.documents:
+            if document.pipeline_version != self.pipeline_version:
+                raise ValueError("document pipeline_version must match manifest pipeline_version")
+        for chunk in self.chunks:
+            document = documents.get(chunk.document_id)
+            if document is None:
+                raise ValueError(f"chunk references unknown document: {chunk.document_id}")
+            if chunk.pipeline_version != self.pipeline_version:
+                raise ValueError("chunk pipeline_version must match manifest pipeline_version")
+            if chunk.source_uri != document.source_uri:
+                raise ValueError("chunk source_uri must match its document provenance")
+            ordinals.setdefault(chunk.document_id, []).append(chunk.ordinal)
+        for document_id, values in ordinals.items():
+            if sorted(values) != list(range(len(values))):
+                raise ValueError(f"chunk ordinals must be unique and contiguous for {document_id}")
         return self
+
+    @staticmethod
+    def _require_unique(field: str, values: list[str]) -> None:
+        if len(values) != len(set(values)):
+            raise ValueError(f"{field} values must be unique")

harness/tht/ports/evidence.py

@@ -1,42 +1,125 @@
-"""Port for discovering and acquiring Evidence source objects.
+"""Credential-free port for discovering and acquiring Evidence objects."""
 
-Source adapters own transport details and credentials.  The values crossing this
-boundary are deliberately credential-free so they can safely become provenance.
-"""
+import re
+from collections.abc import Iterable, Mapping, Sequence
+from datetime import UTC, datetime
+from enum import Enum
+from typing import Protocol, runtime_checkable
+from urllib.parse import parse_qsl, urlsplit
 
-from datetime import datetime
-from typing import Iterable, Protocol, runtime_checkable
-
-from pydantic import BaseModel, ConfigDict, Field, JsonValue, field_validator
+from pydantic import BaseModel, ConfigDict, Field, JsonValue, TypeAdapter, field_validator
 
 
-_SECRET_KEYS = {
-    "api_key",
+class FrozenDict(dict):
+    """A JSON-serializable dict whose mutation operations are disabled."""
+
+    def _immutable(self, *args, **kwargs):
+        raise TypeError("frozen JSON metadata cannot be mutated")
+
+    __delitem__ = _immutable
+    __ior__ = _immutable
+    __setitem__ = _immutable
+    clear = _immutable
+    pop = _immutable
+    popitem = _immutable
+    setdefault = _immutable
+    update = _immutable
+
+
+_CAMEL_BOUNDARY = re.compile(r"(?<=[a-z0-9])(?=[A-Z])")
+_SEPARATORS = re.compile(r"[^a-z0-9]+")
+_NAMESPACED_VALUE = re.compile(r"^[a-z][a-z0-9_-]*:[A-Za-z0-9._:-]+$")
+_CREDENTIAL_KEYS = {
     "apikey",
     "authorization",
+    "authtoken",
+    "bearertoken",
+    "clientsecret",
     "credential",
     "credentials",
     "password",
-    "secret",
-    "token",
+    "passwd",
+    "privatekey",
+    "refreshtoken",
+    "sessioncookie",
+    "xapikey",
+    "accesstoken",
 }
+_JSON_METADATA = TypeAdapter(dict[str, JsonValue])
 
 
-def _reject_credentials(value: JsonValue, path: str = "metadata") -> JsonValue:
-    if isinstance(value, dict):
+def _normalize_key(key: str) -> str:
+    return _SEPARATORS.sub("", _CAMEL_BOUNDARY.sub("_", key).lower())
+
+
+def _is_credential_key(key: str) -> bool:
+    return _normalize_key(key) in _CREDENTIAL_KEYS
+
+
+def _reject_credentials(value, path: str = "metadata") -> None:
+    if isinstance(value, Mapping):
         for key, child in value.items():
-            normalized = key.lower().replace("-", "_")
-            if normalized in _SECRET_KEYS or normalized.endswith(("_password", "_secret", "_token")):
+            if _is_credential_key(str(key)):
                 raise ValueError(f"credential-like metadata key is not allowed: {path}.{key}")
             _reject_credentials(child, f"{path}.{key}")
-    elif isinstance(value, list):
+    elif isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
         for index, child in enumerate(value):
             _reject_credentials(child, f"{path}[{index}]")
+
+
+def freeze_json(value):
+    """Recursively freeze a Pydantic-validated JSON value without changing its JSON shape."""
+    if isinstance(value, Mapping):
+        return FrozenDict({str(key): freeze_json(child) for key, child in value.items()})
+    if isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
+        return tuple(freeze_json(child) for child in value)
+    return value
+
+
+def validate_safe_metadata(value: dict[str, JsonValue]) -> FrozenDict:
+    _reject_credentials(value)
+    return freeze_json(value)
+
+
+def validate_canonical_uri(value: str) -> str:
+    try:
+        parsed = urlsplit(value)
+        _ = parsed.port
+    except ValueError as error:
+        raise ValueError("invalid canonical URI") from error
+    if not parsed.scheme:
+        raise ValueError("canonical URI must include a scheme")
+    if parsed.username is not None or parsed.password is not None:
+        raise ValueError("canonical URI must not contain credentials in userinfo")
+    for key, _ in parse_qsl(parsed.query, keep_blank_values=True):
+        if _is_credential_key(key):
+            raise ValueError("canonical URI must not contain credentials in query parameters")
+    return value
+
+
+def normalize_aware_datetime(value: datetime | None) -> datetime | None:
+    if value is None:
+        return None
+    if value.tzinfo is None or value.utcoffset() is None:
+        raise ValueError("datetime must be timezone-aware")
+    return value.astimezone(UTC)
+
+
+def validate_namespaced_value(value: str) -> str:
+    if not _NAMESPACED_VALUE.fullmatch(value):
+        raise ValueError("value must be namespaced as '<kind>:<stable-value>'")
     return value
 
 
 class _EvidenceValue(BaseModel):
-    model_config = ConfigDict(frozen=True, extra="forbid")
+    model_config = ConfigDict(
+        frozen=True,
+        extra="forbid",
+        revalidate_instances="always",
+        validate_default=True,
+        ser_json_bytes="base64",
+        val_json_bytes="base64",
+    )
 
 
 class SourceObject(_EvidenceValue):
@@ -46,25 +129,48 @@ class SourceObject(_EvidenceValue):
     modified_at: datetime | None = None
     metadata: dict[str, JsonValue] = Field(default_factory=dict)
 
-    @field_validator("metadata")
-    @classmethod
-    def metadata_has_no_credentials(cls, value: dict[str, JsonValue]) -> dict[str, JsonValue]:
-        _reject_credentials(value)
-        return value
+    _source_id = field_validator("source_id")(validate_namespaced_value)
+    _fingerprint = field_validator("fingerprint")(validate_namespaced_value)
+    _safe_uri = field_validator("uri")(validate_canonical_uri)
+    _aware_modified_at = field_validator("modified_at")(normalize_aware_datetime)
+    _frozen_metadata = field_validator("metadata")(validate_safe_metadata)
 
 
 class AcquiredDocument(_EvidenceValue):
+    """Transport result; bytes use explicit base64 encoding in JSON mode."""
+
     source: SourceObject
     content: bytes
     media_type: str | None = None
     acquired_at: datetime | None = None
     metadata: dict[str, JsonValue] = Field(default_factory=dict)
 
-    @field_validator("metadata")
-    @classmethod
-    def metadata_has_no_credentials(cls, value: dict[str, JsonValue]) -> dict[str, JsonValue]:
-        _reject_credentials(value)
-        return value
+    _aware_acquired_at = field_validator("acquired_at")(normalize_aware_datetime)
+    _frozen_metadata = field_validator("metadata")(validate_safe_metadata)
+
+
+class EvidenceSourceErrorCategory(str, Enum):
+    TRANSIENT = "transient"
+    PERMANENT = "permanent"
+
+
+class EvidenceSourceError(Exception):
+    """Classified source failure with credential-free structured diagnostics."""
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        category: EvidenceSourceErrorCategory,
+        details: dict[str, JsonValue] | None = None,
+    ) -> None:
+        super().__init__(message)
+        self.category = EvidenceSourceErrorCategory(category)
+        self.details = validate_safe_metadata(_JSON_METADATA.validate_python(details or {}))
+
+    @property
+    def retryable(self) -> bool:
+        return self.category is EvidenceSourceErrorCategory.TRANSIENT
 
 
 @runtime_checkable