fix: Preserve daemon defaults across explicit tier requests

.github/workflows/ci.yml

@@ -125,10 +125,22 @@ jobs:
         os: [ubuntu-latest, macos-latest, windows-latest]
         python-version: ["3.12"]
         mcp-version: [minimum, latest]
+        pydantic-version: [latest]
         include:
           - os: ubuntu-latest
             python-version: "3.10"
             mcp-version: minimum
+            pydantic-version: latest
+          # Exercise the MCP/Pydantic floor pair and the release that removed
+          # eval_type_backport. Other dependencies still use normal resolution.
+          - os: ubuntu-latest
+            python-version: "3.10"
+            mcp-version: minimum
+            pydantic-version: "2.11.0"
+          - os: ubuntu-latest
+            python-version: "3.12"
+            mcp-version: minimum
+            pydantic-version: "2.14.0"
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - name: Set up Python ${{ matrix.python-version }}
@@ -149,6 +161,7 @@ jobs:
         shell: python
         env:
           MCP_VERSION: ${{ matrix.mcp-version }}
+          PYDANTIC_VERSION: ${{ matrix.pydantic-version }}
         run: |
           import os
           from pathlib import Path
@@ -161,7 +174,9 @@ jobs:
           python = env_dir / ("Scripts/python.exe" if os.name == "nt" else "bin/python")
           requirements = [str(wheel.resolve())]
           if os.environ["MCP_VERSION"] == "minimum":
-              requirements.append("mcp[cli]==1.19.0")
+              requirements.append("mcp[cli]==1.21.1")
+          if os.environ["PYDANTIC_VERSION"] != "latest":
+              requirements.append("pydantic==" + os.environ["PYDANTIC_VERSION"])
           subprocess.run([str(python), "-m", "pip", "install", *requirements], check=True)
           subprocess.run([str(python), "-m", "pip", "check"], check=True)
           subprocess.run([str(python), "-I", "tests/mcp-wheel-smoke.py"], check=True)

.github/workflows/publish.yml

@@ -37,6 +37,12 @@ jobs:
       fail-fast: false
       matrix:
         mcp-version: [minimum, latest]
+        pydantic-version: [latest]
+        include:
+          - mcp-version: minimum
+            pydantic-version: "2.11.0"
+          - mcp-version: minimum
+            pydantic-version: "2.14.0"
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
@@ -54,6 +60,7 @@ jobs:
         shell: python
         env:
           MCP_VERSION: ${{ matrix.mcp-version }}
+          PYDANTIC_VERSION: ${{ matrix.pydantic-version }}
         run: |
           import os
           from pathlib import Path
@@ -66,7 +73,9 @@ jobs:
           python = env_dir / ("Scripts/python.exe" if os.name == "nt" else "bin/python")
           requirements = [str(wheel.resolve())]
           if os.environ["MCP_VERSION"] == "minimum":
-              requirements.append("mcp[cli]==1.19.0")
+              requirements.append("mcp[cli]==1.21.1")
+          if os.environ["PYDANTIC_VERSION"] != "latest":
+              requirements.append("pydantic==" + os.environ["PYDANTIC_VERSION"])
           subprocess.run([str(python), "-m", "pip", "install", *requirements], check=True)
           subprocess.run([str(python), "-m", "pip", "check"], check=True)
           subprocess.run([str(python), "-I", "tests/mcp-wheel-smoke.py"], check=True)

docs/resource-budgets.md

@@ -310,3 +310,38 @@ It does not change the tier worker's adaptive/OOM policy, models, precision
 or retrieval depth. The initial comparison's numerical failures and latency
 regressions remain recorded until separate native validation resolves them.
 Linux/CUDA behavior does not establish MPS memory or thermal performance.
+
+## Server target identity prerequisite (#795)
+
+An explicit embedding request no longer changes the daemon's default model,
+dimension or generation. The same raw tier keeps its loaded snapshot, including
+an empty tier whose default changed while its initial load was in progress.
+Main and fast requests retain that captured identity. A cache miss or a
+different tier request resolves the current selection; changing the global
+default alone does not invalidate an already published empty-tier snapshot.
+`base`, `pro` and `qwen3_256` reuse the same loaded Qwen3 instance; `edge` and
+`model2vec` reuse the same loaded Model2Vec instance. These are the two
+deterministic identities already allowed by the fast lane. Alias updates
+publish one immutable cache snapshot. A failed distinct load preserves the
+previous snapshot.
+
+The removed `qwen3` name still raises the existing migration error, including
+case and whitespace variants. This validation does not mutate the default or
+adopt broader alias normalization from the active-model setter.
+
+Explicit custom and other configuration-dependent cache entries retain their
+existing raw-key reuse policy, including a custom entry that loaded a built-in
+identity or Model2Vec fallback before its configuration changed. They gain no
+new identity or dimension attestation, and the legacy server fallback for `minilm` and `bge-small`
+remains unchanged. Result-size preflight uses the same cache-selection rule
+as loading. Fast-lane inference continues to follow its captured loaded-model
+snapshot, without resolving current global settings.
+
+This correction preserves inference/state lock ownership, request deadlines,
+native batch ceilings and sticky CPU recovery. It does not bound the lifetime
+of separately retained old model references: the previous main instance can
+still be alive while a distinct replacement is constructed, and the optional
+fast CPU instance remains separately owned. A managed preparation lifetime and
+an explicit protocol guarantee for older daemons are subsequent prerequisites.
+No tier-switch caller, local loader, target table or active configuration is
+changed by this checkpoint; issue #795 remains open.

pyproject.toml

@@ -29,8 +29,9 @@ dependencies = [
     "model2vec>=0.6.0,<1.0",
     "sqlite-vec>=0.1.0,<1.0",
     "numpy>=1.24.0,<3.0",
-    # Tool metadata requires 1.19; SDK 2 removes the FastMCP import path.
-    "mcp[cli]>=1.19.0,<2.0",
+    # 1.21.1 removes a private Pydantic import broken by Pydantic 2.14.
+    # SDK 2 removes the FastMCP import path.
+    "mcp[cli]>=1.21.1,<2.0",
     "httpx>=0.27.0,<1.0",
     "sentence-transformers>=3.0.0,<7.0",
     "torch>=2.0.0,<3.0",

tests/mcp-wheel-smoke.py

@@ -182,6 +182,16 @@ async def _check_protocol(root: Path, env: dict[str, str], origin: Path) -> None
 def main() -> None:
     _require(bool(sys.flags.isolated), "Run this probe with python -I")
     distribution = importlib.metadata.distribution("truememory")
+    # Report public package versions before subprocess checks, whose output is
+    # intentionally redacted. Dependency import failures then remain traceable.
+    print(json.dumps({
+        "stage": "dependency_versions",
+        "python_version": ".".join(str(part) for part in sys.version_info[:3]),
+        "truememory_version": distribution.version,
+        "mcp_version": importlib.metadata.version("mcp"),
+        "pydantic_version": importlib.metadata.version("pydantic"),
+        "pydantic_core_version": importlib.metadata.version("pydantic-core"),
+    }), flush=True)
     origin = Path(distribution.locate_file("truememory/__init__.py")).resolve()
     spec = importlib.util.find_spec("truememory")
     _require(origin.is_file() and spec is not None and spec.origin is not None,