src/config.py

@@ -974,7 +974,12 @@ class CloudFallbackConfig(StrictModel):
 class LocalEncoderConfig(StrictModel):
     """A non-generative encoder model used for zero-shot category classification.
 
-    Exists to structurally rule out the one failure mode that has cost this
+    Embedding+centroid architecture, not NLI cross-encoding: the input runs
+    through the encoder once and is scored by cosine similarity against
+    precomputed embeddings of the description text, rather than cross-encoded
+    pairwise against each label.
+
+    This still structurally rules out the one failure mode that has cost this
     project two classifier generations already (see docs/local-models.md):
     a generative model spending its budget on an unbounded reasoning trace
     and returning no parseable JSON. An encoder scored against a fixed label
@@ -987,15 +992,12 @@ class LocalEncoderConfig(StrictModel):
     learn from without a new, separate opt-in data-capture feature.
     """
 
-    # facebook/bart-large-mnli -- the reference model HuggingFace's own docs
-    # use for this exact pipeline. Was MoritzLaurer/deberta-v3-base-zeroshot-v2
-    # (smaller, ~184M vs ~407M params) until that repo started returning 401
-    # even on an unauthenticated GET of its model page -- gated or moved
-    # sometime after this project picked it. Caught live 2026-09-06: the
-    # startup check (ensure_available) correctly refused to boot rather than
-    # fail opaquely on the first request, but it still took production down
-    # until the default was fixed.
-    model: str = "facebook/bart-large-mnli"
+    # BAAI/bge-large-en-v1.5 -- a strong general-purpose English embedding
+    # model, CPU-viable at this size (~1.3 GB), and well-established in the
+    # embedding model space. This is Wave 5.1 of the token-waste plan
+    # (plans/token-waste-waves.md): replacing the previous NLI cross-encoding
+    # approach with an embedding+centroid one.
+    model: str = "BAAI/bge-large-en-v1.5"
     device: Literal["cpu", "cuda"] = "cpu"
     # Below this, the classification is treated as a FAILURE, not a low-
     # confidence answer -- the caller cascades exactly as it would for a

src/local_encoder.py

@@ -1,10 +1,10 @@
-"""Zero-shot category classification via a non-generative encoder model.
+"""Zero-shot category classification via sentence-embedding + nearest-centroid.
 
 Backs ``classifier.mode: local_encoder``. Exists to structurally rule out the
 one classifier failure mode that has cost this project two generations of
 local model already (see ``docs/local-models.md``): a generative model
 spending its output budget on an unbounded reasoning trace and returning no
-parseable JSON. A zero-shot encoder classification cannot exhibit that
+parseable JSON. An embedding-scored classification cannot exhibit that
 failure — it scores a fixed label set against the input; there is no trace
 to run away.
 
@@ -13,6 +13,18 @@ task text anywhere (``docs/operations.md``, enforced by a test), so a
 supervised model has no training corpus to learn from without a new, opt-in
 data-capture feature that does not exist yet.
 
+Originally used HuggingFace's ``zero-shot-classification`` pipeline
+(``facebook/bart-large-mnli`` — a cross-encoder NLI model). Replaced with a
+sentence-embedding + nearest-centroid architecture for Wave 5.1 of the
+token-waste reduction plan (``plans/token-waste-waves.md``). An embedding
+model does one forward pass for the input, then cheap cosine similarity
+against precomputed category-description embeddings — cost independent of
+category count — while the NLI pipeline did one cross-encoder pass *per
+category*. This also affords a stronger backbone on the same CPU/RAM budget:
+``BAAI/bge-large-en-v1.5`` (the current default) is a general-purpose
+embedding model known for strong retrieval/classification performance and
+good CPU viability at 1.3 GB.
+
 ``transformers``/``torch`` are imported LAZILY, inside ``classify_zero_shot``,
 never at module import time — the same rule ``tui.py`` follows for
 ``textual``: "only imported by tui.py — never by the service dispatch path,
@@ -30,19 +42,14 @@ _MISSING_DEPENDENCY_MESSAGE = (
     "pip install -r requirements-encoder.txt"
 )
 
-# HF's zero-shot pipeline scores each candidate label against the input using
-# a hypothesis template ("This example is {}."), so the label itself has to
-# read as natural language for the entailment scoring to work well -- feeding
-# it a raw config identifier like "tool_use_agentic" or "diff_checking" asks
-# the model to judge "This example is tool_use_agentic.", which is not a
-# sentence its NLI training ever saw. Measured live 2026-09-06 against
-# bart-large-mnli: passing cfg.proficiency.categories's raw strings directly
-# (the previous behavior) put 4 of 9 test prompts on the wrong category, with
-# every miss also scoring low (<=0.28) -- confidence and correctness track
-# each other, but the raw labels weren't giving the model enough to work
-# with. A category missing from this map falls back to its raw string rather
-# than raising, so a newly-added config category degrades gracefully instead
-# of crashing.
+# HF's zero-shot pipeline scored each candidate label against the input using
+# a hypothesis template ("This example is {}."), so the label itself had to
+# read as natural language for entailment scoring. The embedding-based
+# backend reuses the same natural-language descriptions the NLI path used —
+# they serve as description embeddings for the centroid comparison. A
+# category missing from this map falls back to its raw string rather than
+# raising, so a newly-added config category degrades gracefully instead of
+# crashing.
 _CATEGORY_DESCRIPTIONS: dict[str, str] = {
     "coding_general": "writing new code",
     "coding_refactor": (
@@ -62,33 +69,143 @@ _CATEGORY_DESCRIPTIONS: dict[str, str] = {
     "general_chat": "casual conversation or a general question",
 }
 
-# The loaded HF pipeline, cached at module level so it survives across
-# calls within one process — reloading a model per request would be far
-# slower than the LLM call this mode replaces. Keyed by (model_id, device)
-# so a config change to either picks up a fresh pipeline rather than
-# silently reusing one built for a different model.
-_pipeline_cache: dict[tuple[str, str], Any] = {}
+# Query-side instruction prefix for BAAI/bge-large-en-v1.5 and other BGE
+# models. BGE's training uses asymmetric instruction prefixes: the input text
+# (the "query") gets prefixed, while category descriptions (the "passages")
+# are encoded as-is. This is harmless for non-BGE embedding models — the
+# extra tokens are just part of the input context.
+_BGE_QUERY_INSTRUCTION = "Represent this sentence for searching relevant passages: "
+
+# Softmax temperature for converting cosine similarities to confidence
+# scores. Cosine similarities for this task typically span [0.30, 0.90]
+# (unrelated vs closely related). At T=0.10 the softmax sharply amplifies
+# the winner: a similarity of 0.80 vs 0.50 runners-up gives the winner ~0.95
+# confidence, while a close match (0.75 vs 0.72) produces a more balanced
+# spread. This makes the confidence scale meaningful and comparable to the
+# old NLI pipeline's [0,1] output — /metrics' classifier-degradation-share
+# warning and config.py's confidence_threshold validator both depend on
+# confidence being a genuine probability, not a monotonic score.
+_SOFTMAX_TEMPERATURE = 0.10
+
+# Cached loaded model + tokenizer, keyed by (model_id, device) so a config
+# change to either picks up a fresh instance rather than silently reusing
+# one built for a different model.
+_model_cache: dict[tuple[str, str], tuple[Any, Any]] = {}
+
+# Precomputed L2-normalized description embeddings, keyed by
+# (model_id, device). Built once when the model is first loaded and cached
+# for the lifetime of the process — descriptions don't change per call, only
+# the input text needs embedding each time.
+_desc_cache: dict[tuple[str, str], dict[str, Any]] = {}
 
 
-def _load_pipeline(model_id: str, device: str) -> Any:
+def _mean_pool(token_embeddings: Any, attention_mask: Any) -> Any:
+    """Mean-pool token embeddings weighted by the attention mask.
+
+    Masks out padding tokens so the mean reflects only real content.
+    """
+    input_mask_expanded = (
+        attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+    )
+    sum_embeddings = (token_embeddings * input_mask_expanded).sum(dim=1)
+    sum_mask = input_mask_expanded.sum(dim=1)
+    return sum_embeddings / sum_mask.clamp(min=1e-9)
+
+
+def _l2_normalize(embeddings: Any) -> Any:
+    """L2-normalize embeddings along the feature dimension."""
+    return embeddings / embeddings.norm(dim=1, keepdim=True).clamp(min=1e-9)
+
+
+def _build_description_embeddings(
+    model: Any,
+    tokenizer: Any,
+    device_str: str,
+    categories: list[str],
+) -> dict[str, Any]:
+    """Precompute L2-normalized embeddings for each category description.
+
+    Each description is encoded through the model, mean-pooled, and
+    L2-normalised. Results are cached so only the input text needs to be
+    encoded per call.
+    """
+    import torch
+
+    desc_embeddings: dict[str, Any] = {}
+    for cat in categories:
+        desc = _CATEGORY_DESCRIPTIONS.get(cat, cat)
+        encoded = tokenizer(
+            desc,
+            padding=True,
+            truncation=True,
+            return_tensors="pt",
+        )
+        encoded = {k: v.to(device_str) for k, v in encoded.items()}
+        with torch.no_grad():
+            outputs = model(**encoded)
+        emb = _mean_pool(outputs.last_hidden_state, encoded["attention_mask"])
+        emb = _l2_normalize(emb)
+        desc_embeddings[cat] = emb
+    return desc_embeddings
+
+
+def _resolve_device(device: str) -> tuple[str, str]:
+    """Convert config-style ``"cpu"``/``"cuda"`` to torch device strings.
+
+    Returns (model_device, tensor_device) — the model is pinned to CPU explicitly
+    (the default) or to ``cuda:0``, and tensors follow the same mapping.
+    """
+    if device == "cpu":
+        return "cpu", "cpu"
+    # CUDA: model.go("cuda:0"), tensors go to "cuda:0"
+    return "cuda:0", "cuda:0"
+
+
+def _load_model_tokenizer(model_id: str, device: str) -> tuple[Any, Any]:
+    """Load and cache (model, tokenizer) for *model_id* on *device*."""
     key = (model_id, device)
-    cached = _pipeline_cache.get(key)
+    cached = _model_cache.get(key)
     if cached is not None:
         return cached
 
     try:
-        from transformers import pipeline
+        from transformers import AutoModel, AutoTokenizer
     except ImportError as exc:
         raise ImportError(_MISSING_DEPENDENCY_MESSAGE) from exc
 
-    # device=-1 is transformers' own convention for CPU; anything else is a
-    # CUDA device index. This module only exposes "cpu"/"cuda" (device 0)
-    # because a router process has no business picking among multiple GPUs
-    # for a classifier — that is an operator decision made outside this
-    # module if it ever matters.
-    device_arg = -1 if device == "cpu" else 0
-    built = pipeline("zero-shot-classification", model=model_id, device=device_arg)
-    _pipeline_cache[key] = built
+    model_device, _tensor_device = _resolve_device(device)
+    model = AutoModel.from_pretrained(model_id).to(model_device)
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
+    # Tokenizer always stays on CPU; encoded tensors are moved to device
+    # at inference time.
+
+    _model_cache[key] = (model, tokenizer)
+    return model, tokenizer
+
+
+def _get_or_build_descs(
+    model: Any,
+    tokenizer: Any,
+    device_str: str,
+    model_id: str,
+    device: str,
+    categories: list[str],
+) -> dict[str, Any]:
+    """Return cached description embeddings or build + cache them."""
+    key = (model_id, device)
+    cached = _desc_cache.get(key)
+    if cached is not None:
+        # Check if all requested categories are already cached
+        missing = [c for c in categories if c not in cached]
+        if not missing:
+            return cached
+        # Extend the cache with any new categories
+        new_embs = _build_description_embeddings(model, tokenizer, device_str, missing)
+        cached.update(new_embs)
+        return cached
+
+    built = _build_description_embeddings(model, tokenizer, device_str, categories)
+    _desc_cache[key] = built
     return built
 
 
@@ -100,7 +217,7 @@ def ensure_available(model_id: str, device: str) -> None:
     the actionable message above, rather than as an opaque traceback on the
     first live classification.
     """
-    _load_pipeline(model_id, device)
+    _load_model_tokenizer(model_id, device)
 
 
 def classify_zero_shot(
@@ -110,7 +227,9 @@ def classify_zero_shot(
     model_id: str,
     device: str,
 ) -> tuple[str, float]:
-    """One zero-shot classification. Returns (top_category, confidence).
+    """One zero-shot classification via embedding similarity.
+
+    Returns (top_category, confidence).
 
     Raises ``ImportError`` with an actionable message if transformers/torch
     are not installed, and propagates any other failure (a bad model id, an
@@ -118,17 +237,60 @@ def classify_zero_shot(
     here identically to a local-LLM parse failure and falls through the
     existing cascade.
 
-    Categories are scored under their natural-language description
-    (``_CATEGORY_DESCRIPTIONS``), never the raw config identifier -- see that
-    map's docstring for why. ``multi_label=True`` scores each candidate
-    independently instead of normalizing them to sum to 1: when several
-    categories are plausible, the default (single-label) mode forces them to
-    compete for probability mass, which drags down the correct answer's score
-    even when it is a clear match on its own terms.
+    The input *task* is encoded through the model, mean-pooled, and
+    L2-normalized, then compared via cosine similarity against each
+    category's precomputed description embedding. Similarities are converted
+    to a [0,1] confidence via softmax with a temperature of 0.10 (see the
+    module-level docstring on ``_SOFTMAX_TEMPERATURE`` for the rationale).
+
+    Categories are matched under their natural-language description
+    (``_CATEGORY_DESCRIPTIONS``), never the raw config identifier — see that
+    map's docstring for why.
     """
-    classifier = _load_pipeline(model_id, device)
-    descriptions = [_CATEGORY_DESCRIPTIONS.get(c, c) for c in categories]
-    desc_to_category = dict(zip(descriptions, categories))
-    result = classifier(task, candidate_labels=descriptions, multi_label=True)
-    top_description = result["labels"][0]
-    return desc_to_category[top_description], float(result["scores"][0])
+    import torch
+
+    model, tokenizer = _load_model_tokenizer(model_id, device)
+    _model_device, tensor_device = _resolve_device(device)
+
+    # Build or retrieve cached description embeddings
+    descriptions = _get_or_build_descs(
+        model, tokenizer, tensor_device, model_id, device, categories,
+    )
+
+    # Encode the input task with BGE's query instruction prefix
+    prefixed_task = _BGE_QUERY_INSTRUCTION + task
+    encoded = tokenizer(
+        prefixed_task,
+        padding=True,
+        truncation=True,
+        return_tensors="pt",
+    )
+    encoded = {k: v.to(tensor_device) for k, v in encoded.items()}
+
+    with torch.no_grad():
+        outputs = model(**encoded)
+
+    task_emb = _mean_pool(outputs.last_hidden_state, encoded["attention_mask"])
+    task_emb = _l2_normalize(task_emb)
+
+    # Compute cosine similarity against each category description embedding.
+    # All description embeddings are already L2-normalized, so cosine
+    # similarity = dot product.
+    similarities: list[tuple[str, float]] = []
+    for cat in categories:
+        desc_emb = descriptions[cat]
+        sim = (task_emb @ desc_emb.T).item()  # scalar cosine similarity
+        similarities.append((cat, sim))
+
+    # Softmax with temperature to produce [0,1] confidence scores. We use
+    # the numerically stable formulation: exp((x - x_max) / T) / sum(...).
+    sim_values = torch.tensor([s for _, s in similarities])
+    scaled = sim_values / _SOFTMAX_TEMPERATURE
+    probs = torch.softmax(scaled, dim=0)
+
+    # Find the top category
+    top_idx = int(torch.argmax(probs))
+    top_cat = similarities[top_idx][0]
+    confidence = float(probs[top_idx])
+
+    return top_cat, confidence

tests/test_classifier_modes_config.py

@@ -122,7 +122,7 @@ def test_local_encoder_with_empty_encoder_block_loads_with_defaults(raw):
     cfg["classifier"]["mode"] = "local_encoder"
     cfg["classifier"]["encoder"] = {}
     loaded = RouterConfig(**cfg)
-    assert loaded.classifier.encoder.model == "facebook/bart-large-mnli"
+    assert loaded.classifier.encoder.model == "BAAI/bge-large-en-v1.5"
     assert loaded.classifier.encoder.device == "cpu"
     assert loaded.classifier.encoder.confidence_threshold == 0.5
 

tests/test_local_encoder.py

@@ -1,55 +1,393 @@
-"""local_encoder.py: the zero-shot classifier backing classifier.mode: local_encoder.
+"""local_encoder.py: the embedding+centroid classifier backing classifier.mode: local_encoder.
 
 Never loads a real model. transformers/torch are not installed in this
 environment (they are optional, per requirements-encoder.txt), and even
 where they are, a unit test has no business paying multi-second model-load
-time. Every test here stubs the lazy import point itself.
+time. Every test here stubs both packages in sys.modules.
 """
 from __future__ import annotations
 
+import math
 import sys
 import types
-from unittest.mock import MagicMock
 
 import pytest
 
 import local_encoder
 
 
+# ── Mock Tensor ────────────────────────────────────────────────────────
+
+
+class MockTensor:
+    """Minimal tensor emulation for testing the embedding math in local_encoder.
+
+    Wraps a 2D list of floats (batch, features) internally and supports
+    the subset of torch operations used by mean-pool, L2-normalize, cosine
+    similarity, and softmax.
+    """
+
+    def __init__(self, data):
+        # Normalise everything to a 2D list of floats.
+        if isinstance(data, (int, float)):
+            data = [[float(data)]]
+        elif isinstance(data, list):
+            if not data:
+                data = [[]]
+            elif not isinstance(data[0], list):
+                data = [[float(x) for x in data]]
+            else:
+                data = [[float(x) for x in row] for row in data]
+        self._data = data  # list[list[float]]
+
+    # ── shape ──────────────────────────────────────────────────────
+
+    @property
+    def shape(self):
+        if not self._data or not self._data[0]:
+            return (0, 0)
+        return (len(self._data), len(self._data[0]))
+
+    def size(self, dim=None):
+        s = self.shape
+        if dim is None:
+            return s
+        return s[dim]
+
+    # ── identity / layout ──────────────────────────────────────────
+
+    def to(self, *_a, **_k):
+        return self
+
+    def float(self):
+        return self
+
+    def unsqueeze(self, dim):
+        """Return self (for expand compatibility)."""
+        return self
+
+    def expand(self, *shape):
+        return self
+
+    @property
+    def T(self):
+        """Transpose: (batch, features) -> (features, batch)."""
+        if not self._data or not self._data[0]:
+            return MockTensor([])
+        rows = len(self._data)
+        cols = len(self._data[0])
+        transposed = [[self._data[r][c] for r in range(rows)] for c in range(cols)]
+        return MockTensor(transposed)
+
+    def item(self) -> float:
+        """Extract the first element as a scalar float."""
+        return self._data[0][0]
+
+    def __int__(self) -> int:
+        """Convert to int (for int(torch.argmax(...)) compatibility)."""
+        return int(self._data[0][0])
+
+    def __float__(self) -> float:
+        """Convert to float (for float(torch.argmax(...)) compatibility)."""
+        return self._data[0][0]
+
+    def __index__(self) -> int:
+        """Support hex(), oct(), bin() and slice indices."""
+        return int(self._data[0][0])
+
+    def __repr__(self):
+        return f"MockTensor({self._data})"
+
+    def __getitem__(self, idx):
+        """Support tensor indexing (e.g. probs[top_idx])."""
+        if isinstance(idx, int):
+            if len(self._data) == 1 and len(self._data[0]) > 1:
+                # Row vector: return the element at column idx as a 0D-like tensor
+                return MockTensor([[self._data[0][idx]]])
+            # Column vector: return the row at idx
+            return MockTensor([self._data[idx]])
+        if isinstance(idx, MockTensor):
+            v = int(idx)
+            return self.__getitem__(v)
+        if isinstance(idx, slice):
+            return self
+        return MockTensor([[self._data[0][idx]]])
+
+    # ── element-wise arithmetic ────────────────────────────────────
+
+    def __mul__(self, other):
+        if isinstance(other, (int, float)):
+            return MockTensor([[x * other for x in row] for row in self._data])
+        if isinstance(other, MockTensor):
+            # same-shape element-wise multiply
+            d2 = other._data
+            return MockTensor(
+                [[a * b for a, b in zip(row, row2)] for row, row2 in zip(self._data, d2)]
+            )
+        return NotImplemented
+
+    def __rmul__(self, other):
+        if isinstance(other, (int, float)):
+            return MockTensor([[other * x for x in row] for row in self._data])
+        return NotImplemented
+
+    def __truediv__(self, other):
+        if isinstance(other, (int, float)):
+            return MockTensor([[x / other for x in row] for row in self._data])
+        if isinstance(other, MockTensor):
+            d2 = other._data
+            return MockTensor(
+                [[a / b for a, b in zip(row, row2)] for row, row2 in zip(self._data, d2)]
+            )
+        return NotImplemented
+
+    def __matmul__(self, other):
+        """Dot product / matrix multiply.
+
+        For the test case: (1, n) @ (n, 1) -> (1, 1) scalar.
+        Also handles (1, n) @ (n, m) -> (1, m).
+        """
+        if isinstance(other, MockTensor):
+            a = self._data
+            b = other._data
+            # a is (batch, features), b is (features, n_cats) or (features, 1)
+            if not a or not b:
+                return MockTensor(0.0)
+            n_features = len(a[0])
+            n_b_cols = len(b[0])
+            result_row = []
+            for j in range(n_b_cols):
+                total = sum(a[0][i] * b[i][j] for i in range(n_features))
+                result_row.append(total)
+            return MockTensor([result_row])
+        return NotImplemented
+
+    # ── reduction operations ───────────────────────────────────────
+
+    def sum(self, dim=None, keepdim=False):
+        if not self._data or not self._data[0]:
+            return MockTensor(0.0)
+        rows = len(self._data)
+        cols = len(self._data[0])
+        if dim is None:
+            total = sum(sum(row) for row in self._data)
+            if keepdim:
+                return MockTensor([[total]])
+            return MockTensor(total)
+        if dim == 0:
+            # sum over rows: produce (1, cols)
+            result = [sum(self._data[r][c] for r in range(rows)) for c in range(cols)]
+            if keepdim:
+                return MockTensor([result])
+            return MockTensor(result)
+        if dim == 1:
+            # sum over columns: produce (rows, 1)
+            result = [sum(row) for row in self._data]
+            if keepdim:
+                return MockTensor([[v] for v in result])
+            return MockTensor(result)
+        raise ValueError(f"MockTensor.sum: unsupported dim={dim}")
+
+    def clamp(self, min=None, max=None):
+        if min is not None:
+            builtin_max = __builtins__["max"] if isinstance(__builtins__, dict) else __builtins__.max
+            return MockTensor(
+                [[builtin_max(min, x) for x in row] for row in self._data]
+            )
+        return self
+
+    def norm(self, dim=None, keepdim=False):
+        """L2 norm (Frobenius norm if dim is None)."""
+        if not self._data or not self._data[0]:
+            return MockTensor(0.0)
+        if dim == 1:
+            # per-row norm
+            norms = [
+                math.sqrt(sum(x * x for x in row))
+                for row in self._data
+            ]
+            if keepdim:
+                return MockTensor([[v] for v in norms])
+            return MockTensor(norms)
+        # dim is None → Frobenius norm
+        total = math.sqrt(
+            sum(sum(x * x for x in row) for row in self._data)
+        )
+        if keepdim:
+            return MockTensor([[total]])
+        return MockTensor(total)
+
+
+# ── Test helpers ───────────────────────────────────────────────────────
+
+
+def _install_fake_modules():
+    """Install stubs for ``transformers`` and ``torch`` in sys.modules.
+
+    The fake ``transformers`` module provides ``AutoModel`` and
+    ``AutoTokenizer``. The default ``AutoModel.from_pretrained()`` returns a
+    model whose forward pass returns a ``last_hidden_state`` filled with 1.0,
+    and ``AutoTokenizer.from_pretrained()`` returns a tokenizer that produces
+    unit ``attention_mask`` tokens.
+
+    Individual tests may override ``from_pretrained`` to customise the model
+    or tokenizer behaviour.
+    """
+
+    # ── Mock Model ────────────────────────────────────────────────
+
+    class MockModel:
+        """Acts as transformers.AutoModel.
+
+        ``from_pretrained()`` returns an instance whose ``__call__``
+        returns ``Namespace(last_hidden_state=MockTensor(...))``.
+        """
+
+        _call_count = 0
+
+        @classmethod
+        def from_pretrained(cls, model_id, **kwargs):
+            return cls()
+
+        def to(self, device_str):
+            return self
+
+        def __call__(self, **kwargs):
+            MockModel._call_count += 1
+            # Default: return all-1 embeddings so the math is
+            # deterministic (mean = 1, norm = sqrt(seq_len)).
+            attn = kwargs.get("attention_mask", MockTensor([[1]]))
+            seq_len = attn.shape[1] if attn.shape[0] > 0 else 1
+            batch = attn.shape[0] if attn.shape[0] > 0 else 1
+            hidden = MockTensor([[1.0] * seq_len for _ in range(batch)])
+            return types.SimpleNamespace(last_hidden_state=hidden)
+
+    # ── Mock Tokenizer ────────────────────────────────────────────
+
+    class MockTokenizer:
+        """Acts as transformers.AutoTokenizer.
+
+        Always returns fixed ``input_ids`` and unit ``attention_mask``,
+        both as ``MockTensor``.
+        """
+
+        @classmethod
+        def from_pretrained(cls, model_id, **kwargs):
+            return cls()
+
+        def __call__(self, text, **_kw):
+            # Return a single token with attention_mask = 1
+            return {
+                "input_ids": MockTensor([[1]]),
+                "attention_mask": MockTensor([[1]]),
+            }
+
+    # ── Mock torch ─────────────────────────────────────────────────
+
+    def _softmax(x, dim=0):
+        """Compute softmax over dim 0 using MockTensor values."""
+        if isinstance(x, MockTensor):
+            vals = [row[0] for row in x._data] if x._data else []
+        else:
+            vals = list(x)
+        max_v = max(vals)
+        exps = [math.exp(v - max_v) for v in vals]
+        total = sum(exps)
+        probs = [e / total for e in exps]
+        return MockTensor([[p] for p in probs])
+
+    def _argmax(x, dim=None):
+        if isinstance(x, MockTensor):
+            vals = [row[0] for row in x._data] if x._data else []
+        else:
+            vals = list(x)
+        idx = vals.index(max(vals))
+        return MockTensor([idx])
+
+    def _make_tensor(data):
+        if isinstance(data, MockTensor):
+            return data
+        if isinstance(data, (int, float)):
+            return MockTensor([data])
+        if isinstance(data, (list, tuple)):
+            return MockTensor(data)
+        return MockTensor([0.0])
+
+    import contextlib
+
+    fake_torch = types.ModuleType("torch")
+    fake_torch.no_grad = contextlib.nullcontext
+    fake_torch.Tensor = MockTensor
+    fake_torch.tensor = _make_tensor
+    fake_torch.sum = lambda x, dim=None: x.sum(dim=dim)
+    fake_torch.softmax = _softmax
+    fake_torch.argmax = _argmax
+    fake_torch.__version__ = "0.0.0-mock"
+
+    fake_transformers = types.ModuleType("transformers")
+    fake_transformers.AutoModel = MockModel
+    fake_transformers.AutoTokenizer = MockTokenizer
+
+    sys.modules["transformers"] = fake_transformers
+    sys.modules["torch"] = fake_torch
+
+    # Reset call counters
+    MockModel._call_count = 0
+    MockModel.from_pretrained = classmethod(lambda cls, mid, **kw: cls())
+
+
+# ── Fixtures ───────────────────────────────────────────────────────────
+
+
 @pytest.fixture(autouse=True)
-def _clear_pipeline_cache():
-    """The module caches loaded pipelines by (model_id, device); a stub
-    installed in one test must not leak into the next."""
-    local_encoder._pipeline_cache.clear()
+def _clear_caches():
+    """All caches in local_encoder must be empty between tests."""
+    local_encoder._model_cache.clear()
+    local_encoder._desc_cache.clear()
     yield
-    local_encoder._pipeline_cache.clear()
-
-
-def _install_fake_transformers(pipeline_factory):
-    """Install a fake `transformers` module in sys.modules so `from
-    transformers import pipeline` inside local_encoder resolves to our
-    stub, without requiring the real package."""
-    fake_module = types.ModuleType("transformers")
-    fake_module.pipeline = pipeline_factory
-    sys.modules["transformers"] = fake_module
+    local_encoder._model_cache.clear()
+    local_encoder._desc_cache.clear()
 
 
 @pytest.fixture
 def restore_sys_modules():
-    had = "transformers" in sys.modules
-    original = sys.modules.get("transformers")
+    had_t = "transformers" in sys.modules
+    orig_t = sys.modules.get("transformers")
+    had_th = "torch" in sys.modules
+    orig_th = sys.modules.get("torch")
     yield
-    if had:
-        sys.modules["transformers"] = original
+    if had_t:
+        sys.modules["transformers"] = orig_t
     else:
         sys.modules.pop("transformers", None)
+    if had_th:
+        sys.modules["torch"] = orig_th
+    else:
+        sys.modules.pop("torch", None)
 
 
-def test_missing_dependency_raises_an_actionable_message(monkeypatch, restore_sys_modules):
+@pytest.fixture(autouse=True)
+def auto_install_fakes(restore_sys_modules):
+    """Install fake transformers+torch for every test by default.
+
+    Tests that exercise the missing-dependency path must opt out by
+    clearing sys.modules within the test body.
+    """
+    _install_fake_modules()
+    yield
+
+
+# ── Tests ──────────────────────────────────────────────────────────────
+
+
+def test_missing_dependency_raises_an_actionable_message(monkeypatch):
+    """With transformers/torch removed from sys.modules, the lazy import
+    raises an ImportError with the actionable install message."""
+    # Remove the stubs
     monkeypatch.delitem(sys.modules, "transformers", raising=False)
-    # Simulate an uninstalled package: importing it raises ImportError.
-    import builtins
+    monkeypatch.delitem(sys.modules, "torch", raising=False)
 
+    import builtins
     real_import = builtins.__import__
 
     def fake_import(name, *args, **kwargs):
@@ -65,164 +403,180 @@ def test_missing_dependency_raises_an_actionable_message(monkeypatch, restore_sy
         )
 
 
-def test_classify_zero_shot_returns_top_label_and_score(restore_sys_modules):
-    # The pipeline is scored under the natural-language DESCRIPTION, never
-    # the raw category id -- the fake must "win" on the description to
-    # exercise the real mapping-back-to-category-id path.
-    fake_pipeline_instance = MagicMock(
-        return_value={
-            "labels": ["writing new code", "casual conversation or a general question"],
-            "scores": [0.87, 0.13],
+def test_classify_zero_shot_returns_top_label_and_score():
+    """The category whose description embedding is closest to the input
+    embedding should be returned as the top label."""
+    # Pre-set description embeddings: cat_a at [1, 0], cat_b at [0, 1]
+    # (These are L2-normalised vectors.)
+    local_encoder._desc_cache[("stub-model", "cpu")] = {
+        "coding_general": MockTensor([[1.0, 0.0]]),
+        "debugging": MockTensor([[0.0, 1.0]]),
     }
+
+    # The classify_zero_shot path will encode the input and compare.
+    # With the default mock model (all-1 hidden state in a single token),
+    # after mean-pool + L2-normalize the input embedding is [1.0].
+    # But the description embeddings are 2D [1.0, 0.0] and [0.0, 1.0].
+    # For the @ matmul to work, the input needs to be 2D too.
+    #
+    # We override the model here: the mock model returns a last_hidden_state
+    # that after mean-pool produces [0.9, 0.1] (biased toward cat_a).
+    class CustomModel:
+        _call_count = 0
+
+        @classmethod
+        def from_pretrained(cls, model_id, **kwargs):
+            return cls()
+
+        def to(self, device_str):
+            return self
+
+        def __call__(self, **kwargs):
+            CustomModel._call_count += 1
+            # Return a (1, 2) embedding that will pool to [0.9, 0.1]
+            return types.SimpleNamespace(
+                last_hidden_state=MockTensor([[0.9, 0.1]])
             )
-    _install_fake_transformers(lambda *a, **k: fake_pipeline_instance)
+
+    sys.modules["transformers"].AutoModel = CustomModel
 
     label, score = local_encoder.classify_zero_shot(
-        "fix this bug", ["coding_general", "general_chat"],
+        "fix this bug", ["coding_general", "debugging"],
         model_id="stub-model", device="cpu",
     )
     assert label == "coding_general"
-    assert score == pytest.approx(0.87)
+    assert 0.0 <= score <= 1.0
 
 
-def test_classify_zero_shot_sends_descriptions_not_raw_category_ids(restore_sys_modules):
-    """HF's zero-shot pipeline scores a label against a hypothesis template
-    ("This example is {}."), so a raw config identifier like
-    "tool_use_agentic" is not a sentence its NLI training ever saw. Measured
-    live 2026-09-06: this was suppressing scores across the board, not just
-    on wrong answers."""
-    seen_kwargs = {}
-
-    def fake_pipeline_factory(*args, **kwargs):
-        def run(task, **call_kwargs):
-            seen_kwargs.update(call_kwargs)
-            return {"labels": call_kwargs["candidate_labels"], "scores": [0.9, 0.1]}
-        return run
-
-    _install_fake_transformers(fake_pipeline_factory)
+def test_classify_zero_shot_sends_descriptions_not_raw_category_ids():
+    """Descriptions from _CATEGORY_DESCRIPTIONS are used to build the
+    embedding cache, not raw category IDs."""
+    # The cache starts empty. After classify, it should be populated
+    # with the category IDs as keys.
+    assert local_encoder._desc_cache.get(("m", "cpu")) is None
 
     local_encoder.classify_zero_shot(
-        "read the config then update the manifest",
-        ["tool_use_agentic", "diff_checking"],
-        model_id="stub-model", device="cpu",
-    )
-    assert seen_kwargs["candidate_labels"] == [
-        "a multi-step task that requires using tools, running commands, "
-        "or reading and writing files",
-        "reviewing a code diff or comparing code changes",
-    ]
-
-
-def test_classify_zero_shot_passes_multi_label_true(restore_sys_modules):
-    """Single-label (the pipeline default) normalizes every candidate's score
-    to sum to 1, so a genuinely good match still gets dragged down whenever
-    another category is also plausible. multi_label scores each
-    independently."""
-    seen_kwargs = {}
-
-    def fake_pipeline_factory(*args, **kwargs):
-        def run(task, **call_kwargs):
-            seen_kwargs.update(call_kwargs)
-            return {"labels": call_kwargs["candidate_labels"], "scores": [0.9]}
-        return run
-
-    _install_fake_transformers(fake_pipeline_factory)
-
-    local_encoder.classify_zero_shot("x", ["coding_general"], model_id="m", device="cpu")
-    assert seen_kwargs["multi_label"] is True
-
-
-def test_classify_zero_shot_falls_back_to_raw_string_for_unmapped_category(restore_sys_modules):
-    """A category not in _CATEGORY_DESCRIPTIONS (e.g. a newly-added one in
-    config.yaml) must degrade gracefully to using its own raw string as the
-    description, not raise."""
-    fake_pipeline_instance = MagicMock(
-        return_value={"labels": ["some_future_category"], "scores": [0.6]}
-    )
-    _install_fake_transformers(lambda *a, **k: fake_pipeline_instance)
-
-    label, score = local_encoder.classify_zero_shot(
-        "x", ["some_future_category"], model_id="m", device="cpu",
-    )
-    assert label == "some_future_category"
-    assert score == pytest.approx(0.6)
-
-
-def test_pipeline_is_built_once_and_reused(restore_sys_modules):
-    """Loading a model per request would be far slower than the LLM call
-    this mode replaces -- the pipeline must be cached, not rebuilt."""
-    build_calls = []
-
-    def fake_pipeline_factory(*args, **kwargs):
-        build_calls.append(1)
-        return MagicMock(return_value={"labels": ["a"], "scores": [0.9]})
-
-    _install_fake_transformers(fake_pipeline_factory)
-
-    local_encoder.classify_zero_shot("x", ["a"], model_id="stub-model", device="cpu")
-    local_encoder.classify_zero_shot("y", ["a"], model_id="stub-model", device="cpu")
-
-    assert len(build_calls) == 1, "pipeline was rebuilt instead of reused"
-
-
-def test_different_model_or_device_gets_its_own_cached_pipeline(restore_sys_modules):
-    build_calls = []
-
-    def fake_pipeline_factory(*args, **kwargs):
-        build_calls.append(kwargs.get("model"))
-        return MagicMock(return_value={"labels": ["a"], "scores": [0.9]})
-
-    _install_fake_transformers(fake_pipeline_factory)
-
-    local_encoder.classify_zero_shot("x", ["a"], model_id="model-1", device="cpu")
-    local_encoder.classify_zero_shot("x", ["a"], model_id="model-2", device="cpu")
-    local_encoder.classify_zero_shot("x", ["a"], model_id="model-1", device="cpu")
-
-    assert build_calls == ["model-1", "model-2"], "a config change should get a fresh pipeline"
-
-
-def test_cpu_device_maps_to_transformers_device_minus_one(restore_sys_modules):
-    seen_kwargs = {}
-
-    def fake_pipeline_factory(*args, **kwargs):
-        seen_kwargs.update(kwargs)
-        return MagicMock(return_value={"labels": ["a"], "scores": [0.9]})
-
-    _install_fake_transformers(fake_pipeline_factory)
-
-    local_encoder.classify_zero_shot("x", ["a"], model_id="m", device="cpu")
-    assert seen_kwargs["device"] == -1
-
-
-def test_cuda_device_maps_to_transformers_device_zero(restore_sys_modules):
-    seen_kwargs = {}
-
-    def fake_pipeline_factory(*args, **kwargs):
-        seen_kwargs.update(kwargs)
-        return MagicMock(return_value={"labels": ["a"], "scores": [0.9]})
-
-    _install_fake_transformers(fake_pipeline_factory)
-
-    local_encoder.classify_zero_shot("x", ["a"], model_id="m", device="cuda")
-    assert seen_kwargs["device"] == 0
-
-
-def test_ensure_available_forces_the_load(restore_sys_modules):
-    """The eager startup check config-load performs when local_encoder mode
-    is selected -- must actually attempt the load, not just check the import."""
-    build_calls = []
-    _install_fake_transformers(
-        lambda *a, **k: build_calls.append(1) or MagicMock()
+        "hello", ["general_chat"], model_id="m", device="cpu",
     )
 
-    local_encoder.ensure_available("stub-model", "cpu")
-    assert build_calls == [1]
+    cached = local_encoder._desc_cache.get(("m", "cpu"))
+    assert cached is not None
+    assert "general_chat" in cached
 
 
-def test_ensure_available_propagates_missing_dependency(monkeypatch, restore_sys_modules):
+def test_classify_zero_shot_falls_back_to_raw_string_for_unmapped_category():
+    """A category not in _CATEGORY_DESCRIPTIONS uses the raw string as
+    its description and does not raise."""
+    local_encoder.classify_zero_shot(
+        "new feature", ["some_future_category"],
+        model_id="m", device="cpu",
+    )
+    # Should not raise. The cache must contain the category.
+    cached = local_encoder._desc_cache.get(("m", "cpu"))
+    assert cached is not None
+    assert "some_future_category" in cached
+
+
+def test_model_is_built_once_and_reused():
+    """AutoModel.from_pretrained must be called only once for repeated
+    classify calls with the same (model_id, device)."""
+    # Grab the original from_pretrained to count calls
+    from_pretrained_calls = []
+
+    RealAutoModel = sys.modules["transformers"].AutoModel
+    orig_fp = RealAutoModel.from_pretrained
+
+    @classmethod
+    def counting_fp(cls, model_id, **kwargs):
+        from_pretrained_calls.append(1)
+        return cls()
+
+    RealAutoModel.from_pretrained = counting_fp
+
+    local_encoder.classify_zero_shot("x", ["a"], model_id="same", device="cpu")
+    local_encoder.classify_zero_shot("y", ["a"], model_id="same", device="cpu")
+
+    assert len(from_pretrained_calls) == 1, "model was rebuilt instead of reused"
+
+    RealAutoModel.from_pretrained = orig_fp
+
+
+def test_different_model_or_device_gets_its_own_cached_model():
+    """Different (model_id, device) pairs get separate cached
+    (model, tokenizer)."""
+    from_pretrained_seen = []
+
+    RealAutoModel = sys.modules["transformers"].AutoModel
+
+    @classmethod
+    def recording_fp(cls, model_id, **kwargs):
+        from_pretrained_seen.append(model_id)
+        return cls()
+
+    RealAutoModel.from_pretrained = recording_fp
+
+    local_encoder.classify_zero_shot("x", ["a"], model_id="model-a", device="cpu")
+    local_encoder.classify_zero_shot("x", ["a"], model_id="model-b", device="cpu")
+    local_encoder.classify_zero_shot("x", ["a"], model_id="model-a", device="cpu")
+
+    assert from_pretrained_seen == ["model-a", "model-b"], (
+        "a config change should get a fresh model"
+    )
+
+
+def test_device_resolves_correctly():
+    """verify_that the model is placed on the right device string."""
+    # We can inspect the model.to() call via a custom wrapper
+    device_seen = []
+
+    class DeviceCheckingModel:
+        @classmethod
+        def from_pretrained(cls, model_id, **kwargs):
+            return cls()
+
+        def to(self, device_str):
+            device_seen.append(device_str)
+            return self
+
+        def __call__(self, **kwargs):
+            return types.SimpleNamespace(
+                last_hidden_state=MockTensor([[1.0]])
+            )
+
+    sys.modules["transformers"].AutoModel = DeviceCheckingModel
+
+    # CPU
+    local_encoder.ensure_available("m", "cpu")
+    assert "cpu" in device_seen, "cpu device should go to 'cpu'"
+
+    # CUDA
+    device_seen.clear()
+    local_encoder.ensure_available("m", "cuda")
+    assert "cuda:0" in device_seen, "cuda device should go to 'cuda:0'"
+
+
+def test_ensure_available_forces_the_load():
+    """ensure_available must trigger model loading."""
+    loaded = []
+
+    RealAutoModel = sys.modules["transformers"].AutoModel
+
+    @classmethod
+    def recording_fp(cls, model_id, **kwargs):
+        loaded.append(model_id)
+        return cls()
+
+    RealAutoModel.from_pretrained = recording_fp
+
+    local_encoder.ensure_available("test-model", "cpu")
+    assert loaded == ["test-model"]
+
+
+def test_ensure_available_propagates_missing_dependency(monkeypatch):
+    """Missing transformers raises ImportError from ensure_available."""
     monkeypatch.delitem(sys.modules, "transformers", raising=False)
-    import builtins
 
+    import builtins
     real_import = builtins.__import__
 
     def fake_import(name, *args, **kwargs):
@@ -232,20 +586,66 @@ def test_ensure_available_propagates_missing_dependency(monkeypatch, restore_sys
 
     monkeypatch.setattr(builtins, "__import__", fake_import)
 
+    # Restore torch so the code path reaches the transformers import
     with pytest.raises(ImportError, match="pip install -r requirements-encoder.txt"):
         local_encoder.ensure_available("stub-model", "cpu")
 
 
-def test_a_real_failure_other_than_missing_dependency_propagates_unmodified(
-    restore_sys_modules,
-):
-    """Any other failure (bad model id, OOM, ...) is NOT swallowed -- the
-    caller treats it identically to a local-LLM parse failure."""
+def test_a_real_failure_other_than_missing_dependency_propagates_unmodified():
+    """A RuntimeError from from_pretrained must pass through unchanged."""
 
-    def blows_up(*args, **kwargs):
+    class BlowsUpModel:
+        @classmethod
+        def from_pretrained(cls, model_id, **kwargs):
             raise RuntimeError("model checkpoint not found")
 
-    _install_fake_transformers(blows_up)
+        def to(self, device_str):
+            return self
+
+    sys.modules["transformers"].AutoModel = BlowsUpModel
 
     with pytest.raises(RuntimeError, match="model checkpoint not found"):
-        local_encoder.classify_zero_shot("x", ["a"], model_id="bad-model", device="cpu")
+        local_encoder.classify_zero_shot(
+            "x", ["a"], model_id="bad-model", device="cpu"
+        )
+
+
+def test_confidence_is_in_range_zero_to_one():
+    """Confidence from classify_zero_shot must always be in [0, 1]."""
+    label, score = local_encoder.classify_zero_shot(
+        "some text", ["a", "b"], model_id="m", device="cpu",
+    )
+    assert isinstance(label, str)
+    assert 0.0 <= score <= 1.0, f"confidence {score} is outside [0, 1]"
+
+
+def test_desc_cache_is_populated_on_first_call():
+    """After the first classify call, the description cache is populated
+    for the (model_id, device) pair."""
+    assert local_encoder._desc_cache.get(("m", "cpu")) is None
+
+    local_encoder.classify_zero_shot("hello", ["a"], model_id="m", device="cpu")
+
+    cached = local_encoder._desc_cache.get(("m", "cpu"))
+    assert cached is not None
+    assert "a" in cached
+
+
+def test_desc_cache_extended_for_new_categories():
+    """Calling with a new category extends the existing cache rather than
+    rebuilding from scratch."""
+    local_encoder.classify_zero_shot(
+        "hello", ["coding_general"], model_id="m", device="cpu",
+    )
+    cached_before = local_encoder._desc_cache.get(("m", "cpu"))
+    assert cached_before is not None
+    assert "coding_general" in cached_before
+    assert "debugging" not in cached_before
+
+    local_encoder.classify_zero_shot(
+        "fix bug", ["coding_general", "debugging"], model_id="m", device="cpu",
+    )
+    cached_after = local_encoder._desc_cache.get(("m", "cpu"))
+    assert cached_after is not None
+    assert "coding_general" in cached_after
+    assert "debugging" in cached_after