fix(s31): close 1 review BLOCKER — EmbedderConfig registry + --embedder-config, never an import path

tests/test_run_cli.py

@@ -10,6 +10,7 @@ content so the dry-run reaches its offline return.
 from __future__ import annotations
 
 import json
+import sys
 from pathlib import Path
 
 import pytest
@@ -604,6 +605,92 @@ def test_semantic_retrieval_with_both_dirs_is_not_refused(tmp_path, capsys) -> N
     assert "--semantic-retrieval" not in out.err
 
 
+def test_embedder_config_valid_file_parses_offline(tmp_path, capsys) -> None:
+    """``--embedder-config <valid>`` is accepted and the run still stops offline — the registry is
+    reachable from the CLI, not merely importable."""
+    cfg = tmp_path / "embedder.json"
+    cfg.write_text('{"type": "fake"}', encoding="utf-8")
+    rc = run.main(
+        [
+            _PID,
+            "--docs-dir",
+            str(BUNDLE_DIR),
+            "--bundle-dir",
+            str(BUNDLE_DIR),
+            "--verdict-dir",
+            str(tmp_path / "inbox"),
+            "--semantic-retrieval",
+            "--embedder-config",
+            str(cfg),
+            "--live-dry-run",
+        ]
+    )
+    assert rc == 0
+    assert "LIVE-DRY-RUN OK" in capsys.readouterr().out
+
+
+def test_embedder_config_missing_file_refuses(tmp_path, capsys) -> None:
+    """Fail-fast startup config: a missing file refuses the run (rc 1, no traceback), surfacing
+    through the existing structured-refusal handler."""
+    rc = run.main(
+        [
+            _PID,
+            "--docs-dir",
+            str(BUNDLE_DIR),
+            "--bundle-dir",
+            str(BUNDLE_DIR),
+            "--verdict-dir",
+            str(tmp_path / "inbox"),
+            "--semantic-retrieval",
+            "--embedder-config",
+            "/nonexistent-embedder-config.json",
+            "--live-dry-run",
+        ]
+    )
+    err = capsys.readouterr().err
+    assert rc == 1
+    assert "refused" in err.lower()
+    assert "Traceback" not in err
+
+
+def test_embedder_config_unknown_type_refuses(tmp_path, capsys) -> None:
+    """A config naming an embedder outside the closed registry is REFUSED, never resolved — this
+    is the CLI-level face of the no-import-path rule."""
+    cfg = tmp_path / "embedder.json"
+    cfg.write_text('{"type": "my_pkg.mod:NetworkEmbedder"}', encoding="utf-8")
+    rc = run.main(
+        [
+            _PID,
+            "--docs-dir",
+            str(BUNDLE_DIR),
+            "--bundle-dir",
+            str(BUNDLE_DIR),
+            "--verdict-dir",
+            str(tmp_path / "inbox"),
+            "--semantic-retrieval",
+            "--embedder-config",
+            str(cfg),
+            "--live-dry-run",
+        ]
+    )
+    err = capsys.readouterr().err
+    assert rc == 1
+    assert "refused" in err.lower()
+    assert "my_pkg" not in sys.modules
+
+
+def test_report_with_embedder_config_is_refused(tmp_path, capsys) -> None:
+    """--report stays an ALLOWLIST: a new config flag must be refused there like every other one,
+    else it would be silently dropped."""
+    cfg = tmp_path / "embedder.json"
+    cfg.write_text('{"type": "fake"}', encoding="utf-8")
+    ledger_file = tmp_path / "ledger.json"
+    SavingsLedger(entries=[]).save(str(ledger_file))
+    rc = run.main(["--report", "--ledger", str(ledger_file), "--embedder-config", str(cfg)])
+    assert rc == 1
+    assert "refused" in capsys.readouterr().err.lower()
+
+
 def test_semantic_retrieval_is_not_refused_in_portfolio_mode(capsys) -> None:
     """The flag is valid in BOTH modes (like --dimension-config), so the portfolio partition must
     not name it. Probed via a run that IS refused for a different flag: the refusal lists

tests/test_semretrieval.py

@@ -21,13 +21,18 @@ from pathlib import Path
 
 import numpy as np
 import pytest
+from pydantic import ValidationError
 
 from portfolio_optimiser.semretrieval import (
     EMBED_DIM,
+    Embedder,
+    EmbedderConfig,
     FakeEmbedder,
     HybridRanker,
     StructuralRetriever,
+    build_embedder,
     cosine,
+    load_embedder_config,
     load_vector_store,
     save_vector_store,
 )
@@ -291,6 +296,61 @@ def test_hybrid_ranker_at_weight_zero_equals_the_structural_ranking() -> None:
     ]
 
 
+# --- The embedder config seam: a CLOSED registry, never an import path -------------------------
+
+
+def test_build_embedder_returns_the_shipped_fake_for_the_known_type() -> None:
+    """The registry's one shipped entry. Mirrors ``tests/test_notify.py``'s build_notifier tests."""
+    built = build_embedder(EmbedderConfig(type="fake"))
+    assert isinstance(built, FakeEmbedder)
+    assert isinstance(built, Embedder)  # satisfies the runtime-checkable protocol
+
+
+def test_build_embedder_rejects_an_unknown_type() -> None:
+    """Closed vocabulary: an unknown type RAISES rather than falling back to a default, and the
+    message names the seam so the constraint is discoverable at the failure."""
+    with pytest.raises(ValueError, match="unknown embedder type"):
+        build_embedder(EmbedderConfig(type="nope"))
+
+
+def test_build_embedder_refuses_an_import_path_instead_of_importing_it() -> None:
+    """SECURITY — the registry must never resolve a dotted ``module:Class`` string from config.
+
+    Doing so would be arbitrary code execution at config-load time and would hand a config file the
+    ability to import a module that opens a socket — routing around this module's no-network guard,
+    which is scoped to ``semretrieval.py`` and blind to a third module by construction. The config
+    is accepted as DATA and then refused; nothing is imported."""
+    hostile = EmbedderConfig(type="evil_pkg.embedders:NetworkEmbedder")
+    with pytest.raises(ValueError, match="unknown embedder type"):
+        build_embedder(hostile)
+    assert "evil_pkg" not in sys.modules
+
+
+def test_embedder_config_rejects_an_empty_type() -> None:
+    with pytest.raises(ValidationError):
+        EmbedderConfig(type="")
+
+
+def test_load_embedder_config_is_fail_fast_on_a_missing_file(tmp_path: Path) -> None:
+    """Authoritative startup config — absence RAISES, in deliberate contrast to the tolerant
+    verdict-inbox RAW layer which skips bad files."""
+    with pytest.raises(FileNotFoundError):
+        load_embedder_config(tmp_path / "nope.json")
+
+
+def test_load_embedder_config_is_fail_fast_on_a_malformed_file(tmp_path: Path) -> None:
+    bad = tmp_path / "embedder.json"
+    bad.write_text('{"type": 17}', encoding="utf-8")
+    with pytest.raises(ValidationError):
+        load_embedder_config(bad)
+
+
+def test_load_embedder_config_round_trips_a_valid_file(tmp_path: Path) -> None:
+    good = tmp_path / "embedder.json"
+    good.write_text('{"type": "fake"}', encoding="utf-8")
+    assert isinstance(build_embedder(load_embedder_config(good)), FakeEmbedder)
+
+
 # --- SC3 / SC7: the vector store is byte-deterministic and fails fast on a row/line mismatch ---