feat(incremental): I-2 digest + consistency-checker (WP-0011 T3)

A Merkle-style digest summarizes the derived tier (per-identity fingerprint +
incident edges as order-independent leaves) so equal states have equal digests
and the digest is stable under equivalent event orders. A ConsistencyChecker
recomputes the authoritative fold from the current source, compares it over a
sampled region, and on mismatch scoped-recomputes just the affected identities —
self-healing missed-delta drift, corrupted internal state, and vanished pages.
Makes derived = f(canonical) verified, not asserted.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/shard_wiki/incremental/__init__.py

@@ -22,6 +22,12 @@ from shard_wiki.incremental.minhash import (
     jaccard,
     shingles,
 )
+from shard_wiki.incremental.verification import (
+    ConsistencyChecker,
+    ConsistencyReport,
+    derived_digest,
+    region_digest,
+)
 
 __all__ = [
     "shingles",
@@ -31,4 +37,8 @@ __all__ = [
     "EquivalenceEdge",
     "EquivalenceIndex",
     "normalized_title",
+    "derived_digest",
+    "region_digest",
+    "ConsistencyReport",
+    "ConsistencyChecker",
 ]

src/shard_wiki/incremental/equivalence.py

@@ -136,6 +136,15 @@ class EquivalenceIndex:
 
     # -- queries -------------------------------------------------------------
 
+    def identities(self) -> frozenset[Identity]:
+        """All identities currently present in the index."""
+        return frozenset(self._entries)
+
+    def fingerprint(self, identity: Identity) -> str | None:
+        """The content fingerprint indexed for ``identity`` (None if absent) — a digest leaf."""
+        entry = self._entries.get(identity)
+        return entry.fingerprint if entry is not None else None
+
     def edges(self) -> frozenset[frozenset[Identity]]:
         """All equivalence edges (content + curator) among currently present identities."""
         present = self._entries.keys()

src/shard_wiki/incremental/verification.py

@@ -0,0 +1,112 @@
+"""I-2 verification — digest + background consistency-checker (SHARD-WP-0011 T3).
+
+``derived = f(canonical)`` is made *verified*, not asserted. A **Merkle-style digest** summarizes
+the derived tier (each identity's content fingerprint + its incident equivalence edges as a leaf,
+order-independently combined into a root) so two derived states are equal iff their digests match.
+A **consistency-checker** recomputes the authoritative fold from the current source, compares it to
+the maintained index over a (sampled) region, and on mismatch performs a **scoped recompute** of
+just the affected identities — self-healing drift from a missed delta or corrupted state.
+
+The digest is a pure function of index state, so it is "maintained alongside deltas" for free and
+is stable under equivalent event orders (leaves are sorted before combination).
+"""
+
+from __future__ import annotations
+
+import hashlib
+from collections.abc import Callable, Iterable
+from dataclasses import dataclass
+
+from shard_wiki.incremental.equivalence import EquivalenceIndex
+from shard_wiki.model import Identity, Page
+
+__all__ = ["region_digest", "derived_digest", "ConsistencyReport", "ConsistencyChecker"]
+
+CuratorEdges = Iterable[tuple[Identity, Identity]]
+
+
+def _leaf(index: EquivalenceIndex, identity: Identity) -> str:
+    """A digest leaf for one identity: its fingerprint + its incident edges (as sorted peers)."""
+    fingerprint = index.fingerprint(identity) or "∅"
+    peers = sorted(
+        str(other)
+        for edge in index.edges()
+        if identity in edge
+        for other in edge
+        if other != identity
+    )
+    payload = f"{identity}|{fingerprint}|{','.join(peers)}"
+    return hashlib.blake2b(payload.encode("utf-8"), digest_size=16).hexdigest()
+
+
+def region_digest(index: EquivalenceIndex, identities: Iterable[Identity]) -> str:
+    """A Merkle-style root over the given identities' leaves (order-independent)."""
+    leaves = sorted(_leaf(index, identity) for identity in identities)
+    root = hashlib.blake2b(digest_size=16)
+    for leaf in leaves:
+        root.update(leaf.encode("utf-8"))
+    return root.hexdigest()
+
+
+def derived_digest(index: EquivalenceIndex) -> str:
+    """The digest of the whole maintained derived tier."""
+    return region_digest(index, index.identities())
+
+
+@dataclass(frozen=True, slots=True)
+class ConsistencyReport:
+    """Outcome of a consistency check: what was examined, whether it drifted, and if it healed."""
+
+    checked: int
+    drifted: bool
+    repaired: bool
+    healthy: bool
+
+
+class ConsistencyChecker:
+    """Compares the maintained index against an authoritative rebuild and repairs drift in place."""
+
+    def __init__(
+        self,
+        index: EquivalenceIndex,
+        pages: Callable[[], Iterable[Page]],
+        curator_edges: Callable[[], CuratorEdges] = lambda: (),
+    ) -> None:
+        self._index = index
+        self._pages = pages
+        self._curator = curator_edges
+
+    def _authoritative(self) -> EquivalenceIndex:
+        expected = EquivalenceIndex(
+            num_bands=self._index.num_bands, threshold=self._index.threshold
+        )
+        expected.build(list(self._pages()), list(self._curator()))
+        return expected
+
+    def check_and_repair(self, sample: Iterable[Identity] | None = None) -> ConsistencyReport:
+        """Verify the (sampled) region against a from-scratch fold; scoped-recompute on mismatch."""
+        source = {p.identity: p for p in self._pages()}
+        expected = self._authoritative()
+        region = (
+            set(sample)
+            if sample is not None
+            else set(source) | set(self._index.identities())
+        )
+
+        drifted = region_digest(self._index, region) != region_digest(expected, region)
+        if not drifted:
+            return ConsistencyReport(len(region), drifted=False, repaired=False, healthy=True)
+
+        self._repair(region, source)
+        healthy = region_digest(self._index, region) == region_digest(expected, region)
+        return ConsistencyReport(len(region), drifted=True, repaired=True, healthy=healthy)
+
+    def _repair(self, region: set[Identity], source: dict[Identity, Page]) -> None:
+        """Scoped recompute: reconcile each affected identity to the current source."""
+        present = self._index.identities()
+        for identity in region:
+            page = source.get(identity)
+            if page is not None:
+                self._index.update(page) if identity in present else self._index.add(page)
+            elif identity in present:
+                self._index.remove(identity)