feat(evidence): evaluate retrieval with a small fixture

harness/tht/cli/evidence_cmd.py

@@ -10,6 +10,7 @@ from typing import Annotated
 
 import typer
 
+from tht.cli.config_cmd import CONFIG_OPT
 from tht.evidence import (
     EvidencePreparationError,
     PiEvidenceRestructurer,
@@ -63,6 +64,48 @@ def _findings_payload(findings) -> list[dict[str, object]]:
     return [finding.__dict__ for finding in findings]
 
 
+def evaluate_from_config(
+    workspace_root: Path,
+    config: Path,
+    *,
+    generation: str | None = None,
+) -> dict[str, object]:
+    """Evaluate one stored Evidence generation without changing corpus or vectors."""
+    from tht.adapters.factory import build_vector_store
+    from tht.cli.schema_cmd import _load_config_or_exit
+    from tht.cli.vector_cmd import make_embedder
+    from tht.evidence.canonical import load_curated_tree
+    from tht.evidence.corpus.store import CorpusStore
+    from tht.evidence.evaluation import evaluate_retrieval, load_evaluation_fixture
+
+    cfg = _load_config_or_exit(config)
+    store = CorpusStore(cfg.paths.artifacts.parent / "corpus")
+    manifest = store.manifest(generation) if generation is not None else store.active_manifest()
+    if manifest is None:
+        raise RuntimeError("active Evidence generation is unavailable")
+    document_generations = manifest.metadata.get("document_generations")
+    if not isinstance(document_generations, dict):
+        raise TypeError("Evidence generation is invalid")
+    language = {"en": "english", "it": "italian"}.get(cfg.language)
+    if language is None:
+        raise RuntimeError("workspace language is unsupported for Qdrant BM25")
+    report = evaluate_retrieval(
+        load_evaluation_fixture(workspace_root / "evidence" / "evaluation.yaml"),
+        workspace_revision=cfg._workspace_revision,
+        vector_generation=manifest.vector_generation,
+        document_generations=document_generations,
+        workspace_id=cfg._workspace_id,
+        language=language,
+        searcher=build_vector_store(cfg),
+        embedder=make_embedder(cfg.embeddings),
+        expected_kinds={
+            evidence.id: evidence.kind
+            for evidence in load_curated_tree(workspace_root / "evidence" / "curated")
+        },
+    )
+    return report.model_dump()
+
+
 @evidence_app.command("prepare")
 def prepare_cmd(
     workspace_root: Path,
@@ -106,6 +149,36 @@ def validate_cmd(
     raise typer.Exit(code=1)
 
 
+@evidence_app.command("evaluate")
+def evaluate_cmd(
+    workspace_root: Path,
+    config: Path = CONFIG_OPT,
+    generation: str | None = typer.Option(None, "--generation"),
+    json_output: bool = typer.Option(False, "--json"),
+) -> None:
+    """Evaluate active or selected Evidence retrieval generation without publishing it."""
+    root = _canonical_worktree(workspace_root)
+    try:
+        payload = evaluate_from_config(root, config, generation=generation)
+    except Exception:  # noqa: BLE001 - CLI reports a safe operational failure.
+        _emit({
+            "schemaVersion": 1,
+            "operation": "evidence_evaluate",
+            "status": "failed",
+            "code": "evaluation_failed",
+        }, json_output)
+        raise typer.Exit(code=1) from None
+    payload = {
+        "schemaVersion": 1,
+        "operation": "evidence_evaluate",
+        "status": "passed" if payload["passed"] else "failed",
+        **payload,
+    }
+    _emit(payload, json_output)
+    if not payload["passed"]:
+        raise typer.Exit(code=1)
+
+
 @evidence_app.command("resolve")
 def resolve_cmd(
     workspace_root: Path,