feat(evidence): evaluate retrieval with a small fixture

harness/tht/evidence/corpus/pipeline.py

@@ -7,7 +7,7 @@ import json
 import logging
 import re
 import uuid
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence
 from dataclasses import asdict, dataclass, field
 from datetime import UTC
 from pathlib import Path
@@ -127,6 +127,7 @@ class CorpusPipeline:
         vector_store: VectorStore, embedding_model: str, embedding_dimensions: int,
         chunk_policy: ChunkPolicy, pipeline_version: str, retain_published_generations: int = 3,
         workspace_id: str | None = None, sparse_language: str = "italian",
+        candidate_evaluator: Callable[[CorpusManifest], object] | None = None,
     ) -> None:
         self.store = store
         self.sources = sources
@@ -143,6 +144,14 @@ class CorpusPipeline:
         if sparse_language not in {"english", "italian"}:
             raise ValueError("unsupported Qdrant BM25 language")
         self.sparse_language = sparse_language
+        self.candidate_evaluator = candidate_evaluator
+
+    def _evaluate_candidate(self, manifest: CorpusManifest) -> None:
+        if self.candidate_evaluator is None:
+            return
+        report = self.candidate_evaluator(manifest)
+        if getattr(report, "passed", False) is not True:
+            raise PipelineError("candidate retrieval evaluation failed")
 
     def _assert_workspace_binding(self) -> None:
         manifest = self.store.active_manifest()
@@ -658,6 +667,7 @@ class CorpusPipeline:
                     raise
             generation = read(context, "plan.json")["generation"]
             try:
+                self._evaluate_candidate(CorpusManifest.model_validate(read(context, "manifest.json")))
                 self.store.publish(generation)
             except Exception:
                 compensate(context)
@@ -791,6 +801,7 @@ class CorpusPipeline:
                     manifest, {document.document_id: document.content for document in documents},
                     generation=generation,
                 )
+            self._evaluate_candidate(manifest)
             self.store.publish(staged)
             self.gc(workspace_root=self.store.root.parent)
         except AtomicContentTooLargeError as error: