feat: unify installation model catalog

harness/README.md

@@ -29,8 +29,9 @@ Keys are never logged; URLs are fine. Rotate any key that appeared in chat.
 
 ### `workspaces/<name>.yaml`
 
-A schema-v3 workspace wires the external relational DWH to one internal Qdrant collection
-and the internal Ollama embedding model. Evidence paths remain workspace-local, while Qdrant
+A schema-v4 workspace wires the external relational DWH to one internal Qdrant collection.
+The embedding identity and dimensions come from the Installation Model Catalog, not this
+workspace descriptor. Evidence paths remain workspace-local, while Qdrant
 stores the derived semantic projection for schema, Evidence, Memory, and solved-question
 records. The legacy files under `workspaces/` are retained as migration fixtures; new
 operator-facing descriptors live in the workspace Git registry. `${THT_*}` tokens expand

harness/tests/test_evidence_pi_restructurer.py

@@ -36,8 +36,7 @@ def _candidate():
 def test_pi_restructurer_uses_an_ephemeral_no_tools_invocation(tmp_path, monkeypatch):
     from tht.evidence import authoring
 
-    monkeypatch.setenv("PI_PROVIDER", "zai")
-    monkeypatch.setenv("PI_MODEL", "glm-5.2")
+    monkeypatch.setenv("THT_DEFAULT_SESSION_MODEL", "zai/glm-5.2")
     monkeypatch.setenv("PI_THINKING", "medium")
     calls = []
 

harness/tests/test_local_compose_contract.py

@@ -34,8 +34,13 @@ def test_local_compose_uses_the_generic_external_endpoint_contract():
         assert name in environment
     assert environment["THT_INTERNAL_QDRANT_URL"] == "http://qdrant:6333"
     assert environment["THT_INTERNAL_EMBEDDING_URL"] == "http://embedding:11434"
-    assert environment["THT_INTERNAL_EMBEDDING_MODEL"] == "qwen3-embedding:0.6b"
-    assert environment["THT_INTERNAL_EMBEDDING_DIMENSIONS"] == "1024"
+    # Identity and dimensions come only from the installation-generated Compose projection.
+    for name in (
+        "THT_INTERNAL_EMBEDDING_ID",
+        "THT_INTERNAL_EMBEDDING_MODEL",
+        "THT_INTERNAL_EMBEDDING_DIMENSIONS",
+    ):
+        assert name not in environment
     for name in ("THT_VEC_REST_URL", "THT_VEC_WRITE_REST_URL", "THT_OLLAMA_URL"):
         assert name not in environment
 
@@ -53,6 +58,7 @@ def test_local_compose_uses_the_generic_external_endpoint_contract():
 
     model_init = compose["services"]["embedding-model-init"]
     assert "ports" not in model_init
+    assert "OLLAMA_MODEL" not in model_init.get("environment", {})
     assert model_init["depends_on"]["embedding"]["condition"] == "service_healthy"
     assert compose["services"]["core"]["depends_on"]["embedding-model-init"]["condition"] == "service_completed_successfully"
 

harness/tests/test_qdrant_cli_commands.py

@@ -6,6 +6,7 @@ from datetime import UTC, datetime
 from pathlib import Path
 from types import SimpleNamespace
 
+import pytest
 from typer.testing import CliRunner
 
 from tht.cli import app
@@ -13,6 +14,12 @@ from tht.memory import MemoryRecord, save_registry
 from tht.ports.vector import VectorStoreError
 
 
+@pytest.fixture(autouse=True)
+def _ignore_operator_profile(monkeypatch):
+    """Exercise the self-contained server runtime fixture, not the developer's harness/.env."""
+    monkeypatch.delenv("THT_PROFILE", raising=False)
+
+
 class _FakeEmbedder:
     def embed_documents(self, documents):
         return [[0.1] * 4 for _ in documents]

harness/tht/evidence/authoring.py

@@ -228,14 +228,15 @@ class PiEvidenceRestructurer:
                 "--no-skills",
                 "--extension", str(self._json_mode_extension),
             ]
-            for option, environment_name in (
-                ("--provider", "PI_PROVIDER"),
-                ("--model", "PI_MODEL"),
-                ("--thinking", "PI_THINKING"),
-            ):
-                value = os.environ.get(environment_name)
-                if value:
-                    argv.extend((option, value))
+            canonical_model = os.environ.get("THT_DEFAULT_SESSION_MODEL")
+            if canonical_model:
+                provider, separator, model = canonical_model.partition("/")
+                if separator != "/" or not provider or not model:
+                    raise EvidencePreparationError("pi_restructure_failed", request.source_file)
+                argv.extend(("--provider", provider, "--model", model))
+            thinking = os.environ.get("PI_THINKING")
+            if thinking:
+                argv.extend(("--thinking", thinking))
             argv.extend((
                 "--system-prompt", system_prompt,
                 f"@{request_path}",