CLAUDE.md

@@ -1025,6 +1025,32 @@ The items below remain open.
     `POST /outcome` is the answer for streamed traffic: it arrives afterwards,
     so it works identically either way.
 
+5. **`local_encoder` noise isolation — built for the confirmed shapes; residuals below.**
+   The raw task's fenced code blocks, `Tool result:`-shaped lines, and closed
+   `<system-reminder>` spans are now stripped by `_isolate_task_text` (pure,
+   stdlib-only, deterministic) before the fit + embed pass — `classify_zero_shot`
+   runs isolate → fit → prefix, so cleaning happens first and a noisy task often
+   fits the token window outright. Measured on the real `BAAI/bge-large-en-v1.5`
+   (2026-09-19): 8 clean one-sentence tasks scored 8/8, but the same instructions
+   wrapped in that noise scored 2/8 with the truncation fix already in place —
+   and tail-biased windowing ALONE also scored 2/8 on long noisy pairs, so the
+   fit does not subsume isolation. Post-isolation: 8/8 on short and long noisy
+   pairs, clean-vs-noisy pair consistency 8/8 + 8/8; `'[code]'`/`'[elided]'`
+   placeholder tokens measured worse than pure removal (7/8, 6/8 on long pairs)
+   and were rejected; a 20%-ratio floor guard measured harmful (3/8 — it reverts
+   exactly the short noisy inputs isolation exists to fix) in favor of an
+   absolute 24-char floor that only catches near-all-code inputs. Offline
+   regression: a noise tripwire (same instruction bare vs wrapped must classify
+   identically) fails against pre-isolation code and passes after. Still not
+   built: attention-masking de-weighting as an alternative to stripping (it
+   would preserve the noise tokens' presence without letting them dominate);
+   unclosed `<system-reminder>` tags and un-fenced diff hunks are left in place
+   (only closed-tag spans, fenced blocks, and marker-prefixed lines are
+   stripped); and a code-grounded instruction whose pasted snippet is the
+   subject can still land on a near-category — measured 3/4 on a 4-task
+   grounded set, the residual miss being description similarity
+   ("refactor this helper" + code → `debugging`), not noise dominance.
+
 ## Gaming mode, and the backoff that used to do nothing
 
 **The classifier circuit breaker did not break the circuit.**

src/local_encoder.py

@@ -34,7 +34,10 @@ so the router itself has no UI dependency." A deployment that never selects
 
 from __future__ import annotations
 
-from typing import Any, Optional
+import re
+from typing import Any, Final
+
+import logs
 
 _MISSING_DEPENDENCY_MESSAGE = (
     "classifier.mode is 'local_encoder' but the 'transformers'/'torch' "
@@ -87,6 +90,228 @@ _BGE_QUERY_INSTRUCTION = "Represent this sentence for searching relevant passage
 # confidence being a genuine probability, not a monotonic score.
 _SOFTMAX_TEMPERATURE = 0.10
 
+# ── Task-text isolation (structural agent-session noise) ──────────────
+#
+# Production task text is raw user_content (dispatcher passes the RAW
+# task, not the clamped/framed user_content): fenced code blocks, tool
+# call/result stanzas, and system-reminder-style wrappers surround the
+# prose instruction. Mean pooling weights every token equally, so when
+# noise tokens outnumber the instruction words the pooled embedding —
+# and therefore the verdict — describes the noise, not the intent.
+# Observed on the real BAAI/bge-large-en-v1.5 (2026-09-19): 8 clean
+# one-sentence tasks scored 8/8 correct, but the same instructions
+# wrapped in a fenced code block + a "Tool result:" line + a
+# <system-reminder> tag classified 2/8 — every non-summarization
+# instruction flipped to a file/code-flavored category. Tail-biased
+# windowing (the fit below) does NOT fix this: measured 2/8 on long
+# noisy inputs too, because the window keeps code and instruction
+# mixed. The helper below strips that structural noise BEFORE the fit +
+# embed pass; cleaning first also means a noisy task often fits the
+# window outright, so the fit sees prose, not a code wall.
+#
+# Deliberately conservative, pure, stdlib-only (``re``), deterministic,
+# and free of false strip targets for ordinary prose: inline code spans
+# (``like this``) are KEPT — instructions reference real file names and
+# commands, which are signal, not noise.
+
+# Fenced code blocks: strip the content outright. Empirically decided
+# placeholder vs pure removal on the real model (2026-09-19, 8 noisy
+# pairs x short+long): removal scored 8/8 on both, while '[code]' and
+# '[elided]' placeholders scored 7/8 and 6/8 on long pairs — the
+# placeholder token itself pulls the pooled embedding toward code-
+# flavored categories and costs accuracy, so the winner is removal.
+_FENCE_PAIR_RE: Final = re.compile(r"```.*?```", re.DOTALL)
+_FENCE_OPEN_TAIL_RE: Final = re.compile(r"```.*\Z", re.DOTALL)
+
+# System-reminder-style wrapper tags: agent harnesses inject session
+# metadata inside custom tags; the content is context, not the task.
+# Closed-tag spans only — an unclosed tag is left alone rather than
+# guessing where it ends (the floor guard still protects the verdict).
+_ISOLATE_WRAPPER_TAGS: Final[tuple[str, ...]] = ("system-reminder",)
+_WRAPPER_TAG_RE: Final = re.compile(
+    r"<({})\b[^>]*>.*?</\s*\1\s*>".format("|".join(_ISOLATE_WRAPPER_TAGS)),
+    re.DOTALL | re.IGNORECASE,
+)
+
+# Tool-result lines: "Tool result: file read successfully, 12 lines" and
+# the call/output variants seen in agent-session transcripts. Line-level
+# and conservative: the line must START with the marker (case-
+# insensitive), so prose merely mentioning a tool result survives.
+_TOOL_LINE_RE: Final = re.compile(
+    r"^[ \t]*tool[\s_-]*(?:result|output|call)[\s_-]*:[^\n]*\n?",
+    re.IGNORECASE | re.MULTILINE,
+)
+
+# Floor guard: if stripping leaves (almost) nothing, the input was
+# itself one big code block — e.g. a bare file dump with no instruction.
+# Embedding a bare placeholder would score a meaningless stub, so fall
+# back to the original text, whose code content is then the only signal
+# available. Measured (2026-09-19, real model): code-GROUNDED tasks
+# (instruction + pasted code) classify BETTER stripped even though
+# stripping keeps well under 20% of the characters, so the guard keys on
+# an absolute character floor only — a ratio guard would revert exactly
+# the noise-wrapped short inputs the isolation exists to fix.
+_ISOLATE_MIN_CHARS: Final[int] = 24
+
+
+def _isolate_task_text(task: str) -> str:
+    """Strip structural agent-session noise from *task* before embedding.
+
+    Removes fenced code block bodies, system-reminder-style wrapper tag
+    spans, and tool call/result lines — the shapes that dominate a
+    mean-pooled embedding when they outnumber the prose instruction.
+    Inline code spans and ordinary prose are kept. Pure string in,
+    string out; no I/O, no state.
+
+    If stripping leaves fewer than _ISOLATE_MIN_CHARS — the input was
+    essentially only noise (a bare code dump, no instruction) — the
+    original text is returned: the code content is then the only signal
+    there is, and a stripped stub would be a worse embedding target.
+
+    Idempotent: already-isolated text contains none of the stripped
+    shapes, so a second pass returns it unchanged.
+    """
+    if not task:
+        return task
+    text = _FENCE_PAIR_RE.sub("", task)
+    text = _FENCE_OPEN_TAIL_RE.sub("", text)
+    text = _WRAPPER_TAG_RE.sub("", text)
+    text = _TOOL_LINE_RE.sub("", text)
+    if len(text) < _ISOLATE_MIN_CHARS:
+        return task
+    return text
+
+
+# ── Token-budget fitting (pre-tokenization window selection) ──────────
+#
+# The final ``tokenizer(...)`` call in ``classify_zero_shot`` passes
+# ``truncation=True`` with no ``max_length``, so HuggingFace truncates to
+# the tokenizer's ``model_max_length`` — 512 tokens for the BERT-family
+# BGE models — and, with the default ``truncation_side='right'``, KEEPS
+# the head. The head of an agent-session prompt is thousands of chars of
+# shared boilerplate while the task-specific content lives in the tail,
+# so every long request scored the same truncated prefix and returned the
+# same verdict (observed live on 2026-09-18: six distinct 6.4k-6.6k-char
+# requests, all classified ``diff_checking`` at confidence 0.192363).
+# The helpers below fit the task to the real token window BEFORE that
+# final tokenize, choosing ourselves which tokens survive.
+
+# Tail-biased head+tail split: when the task overflows the budget, the
+# head slice keeps this share of the content budget and the tail gets the
+# rest. A head slice survives at all because the raw task text can carry
+# its instruction at either end ("Translate this: <doc>" vs "<doc> —
+# translate this") — the same rationale as dispatcher.clamp_for_classifier,
+# and it applies here directly because the local_encoder path receives the
+# RAW task (dispatcher._classify_via_configured_mode), not the clamped/
+# framed user_content the chat modes see. The head share stays small
+# because the live failure mode was the opposite collapse: shared head
+# boilerplate crowding out the tail.
+_FIT_HEAD_FRACTION: Final[float] = 0.25
+
+# Marker spliced between the kept head and tail, mirroring
+# dispatcher.clamp_for_classifier's elision convention. Its own token
+# cost is measured per tokenizer and reserved from the budget.
+_FIT_ELISION_MARKER: Final[str] = "\n\n[... content elided ...]\n\n"
+
+# decode() -> re-tokenize is not guaranteed token-identical for every
+# input (unicode normalization, odd graphemes), so the fitted text is
+# built a few tokens UNDER budget. If drift ever exceeds the margin, the
+# final ``truncation=True`` clips a token or two off the tail — a
+# bounded, graceful degradation, never a crash.
+_FIT_SAFETY_MARGIN_TOKENS: Final[int] = 8
+
+# model_max_length values some tokenizers report when the true window is
+# unknown (HF's VERY_LARGE_INTEGER is int(1e30)). Unusable for budget
+# math; falls back to the documented constant below.
+_MODEL_MAX_LENGTH_SENTINEL: Final[int] = int(1e30)
+
+# Conservative window used when model_max_length is missing or a
+# sentinel: 512 is the BERT-family window the BGE models actually have.
+_FALLBACK_MODEL_MAX_LENGTH: Final[int] = 512
+
+# Set once the fallback above has been logged, so a misconfigured
+# tokenizer logs one warning instead of one per request.
+_budget_fallback_logged = False
+
+
+def _special_tokens_allowance(tokenizer: Any) -> int:
+    """Tokens the final tokenize adds around the text ([CLS]/[SEP] for
+    BERT-family tokenizers, including the BGE models)."""
+    add_fn = getattr(tokenizer, "num_special_tokens_to_add", None)
+    if add_fn is None:
+        return 2
+    return max(0, int(add_fn()))
+
+
+def _usable_token_budget(tokenizer: Any) -> int:
+    """Tokens available for task content under the model's real window.
+
+    usable = model_max_length - len(BGE instruction prefix tokens)
+             - special tokens added by the final tokenize.
+
+    A missing or sentinel ``model_max_length`` falls back to
+    _FALLBACK_MODEL_MAX_LENGTH (logged once) rather than raising — an
+    unusual tokenizer must degrade conservatively, not kill the request.
+    """
+    global _budget_fallback_logged
+    max_len = getattr(tokenizer, "model_max_length", None)
+    if (
+        not isinstance(max_len, int)
+        or max_len <= 0
+        or max_len >= _MODEL_MAX_LENGTH_SENTINEL
+    ):
+        if not _budget_fallback_logged:
+            _budget_fallback_logged = True
+            logs.warning(
+                "encoder_token_budget_fallback",
+                reported_model_max_length=str(max_len),
+                fallback=_FALLBACK_MODEL_MAX_LENGTH,
+            )
+        max_len = _FALLBACK_MODEL_MAX_LENGTH
+    prefix_ids = tokenizer(
+        _BGE_QUERY_INSTRUCTION, add_special_tokens=False,
+    )["input_ids"]
+    return max_len - len(list(prefix_ids)) - _special_tokens_allowance(tokenizer)
+
+
+def _fit_task_to_token_budget(task: str, tokenizer: Any) -> str:
+    """Fit *task* into the model's token window, biased toward the tail.
+
+    Tokenizes the bare task (no special tokens, no truncation) and, when
+    it overflows the usable budget, keeps a small head slice plus a
+    tail-majority window (see _FIT_HEAD_FRACTION for why a head slice
+    survives at all), spliced with the elision marker. The window is
+    selected on real token ids and decoded back to text — no
+    chars-per-token arithmetic anywhere.
+
+    Returns *task* unchanged when it already fits (including the empty
+    string). Never mutates tokenizer state — the cached tokenizer is
+    shared across requests — and the final ``truncation=True`` in
+    classify_zero_shot stays as a last-resort backstop.
+    """
+    budget = _usable_token_budget(tokenizer)
+    if budget <= 0:
+        # Pathological window (smaller than prefix + specials): leave the
+        # task alone; the final tokenize does whatever it can.
+        return task
+    ids = list(tokenizer(task, add_special_tokens=False)["input_ids"])
+    if len(ids) <= budget:
+        return task
+    marker_tokens = len(list(
+        tokenizer(_FIT_ELISION_MARKER, add_special_tokens=False)["input_ids"]
+    ))
+    content_budget = budget - marker_tokens - _FIT_SAFETY_MARGIN_TOKENS
+    if content_budget <= 0:
+        # Degenerate window: no room for the marker or the margin — keep
+        # a pure tail, which is where task-specific content lives.
+        return tokenizer.decode(ids[-budget:], skip_special_tokens=True)
+    head_n = int(content_budget * _FIT_HEAD_FRACTION)
+    tail_n = content_budget - head_n
+    head_text = tokenizer.decode(ids[:head_n], skip_special_tokens=True)
+    tail_text = tokenizer.decode(ids[-tail_n:], skip_special_tokens=True)
+    return head_text + _FIT_ELISION_MARKER + tail_text
+
+
 # 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.
@@ -246,6 +471,15 @@ def classify_zero_shot(
     Categories are matched under their natural-language description
     (``_CATEGORY_DESCRIPTIONS``), never the raw config identifier — see that
     map's docstring for why.
+
+    Long inputs are first fitted to the model's real token window (see
+    ``_fit_task_to_token_budget``): the tokenizer's own silent truncation
+    would otherwise keep only the shared head of a long prompt and drop
+    the task-specific tail. Before that, structural agent-session noise
+    (fenced code blocks, tool call/result lines, system-reminder-style
+    tags) is stripped (see ``_isolate_task_text``): mean pooling weights
+    every token equally, so unwrapped noise outnumbers the instruction
+    and the verdict scores the noise, not the intent.
     """
     import torch
 
@@ -257,8 +491,16 @@ def classify_zero_shot(
         model, tokenizer, tensor_device, model_id, device, categories,
     )
 
-    # Encode the input task with BGE's query instruction prefix
+    # Isolate then fit, in that order: stripping the noise first means
-    prefixed_task = _BGE_QUERY_INSTRUCTION + task
+    # the window sees prose and a noisy task often fits outright; the
+    # fit then protects whatever long prose survives. Without isolation
+    # the tokenizer below silently right-truncates to model_max_length
+    # (512 for the BGE models), which for long agent-session prompts
+    # scores only the shared head boilerplate and discards the
+    # task-specific tail: every long request returns the same verdict.
+    isolated_task = _isolate_task_text(task)
+    fitted_task = _fit_task_to_token_budget(isolated_task, tokenizer)
+    prefixed_task = _BGE_QUERY_INSTRUCTION + fitted_task
     encoded = tokenizer(
         prefixed_task,
         padding=True,

tests/test_local_encoder.py

@@ -267,29 +267,64 @@ def _install_fake_modules():
     class MockTokenizer:
         """Acts as transformers.AutoTokenizer.
 
-        Always returns fixed ``input_ids`` and unit ``attention_mask``,
+        Emulates the parts of the HF contract the encoder path relies
-        both as ``MockTensor``.
+        on: a known ``model_max_length`` (512, BERT/BGE family),
+        right-side truncation when called with ``truncation=True`` and
+        no ``max_length`` (HF's default ``truncation_side='right'``),
+        and plain Python id lists when called without ``return_tensors``
+        (the real BatchEncoding behaves the same). Ids are word lengths,
+        so token counts stay content-dependent.
         """
 
+        model_max_length = 512
+        truncation_side = "right"
+
         @classmethod
         def from_pretrained(cls, model_id, **kwargs):
             return cls()
 
-        def __call__(self, text, **_kw):
+        def __call__(
-            # Return a single token with attention_mask = 1
+            self, text, add_special_tokens=True, truncation=False,
+            padding=True, return_tensors=None, max_length=None, **_kw,
+        ):
+            ids = [len(word) for word in str(text).split()]
+            if truncation:
+                keep = self.model_max_length - (2 if add_special_tokens else 0)
+                ids = ids[:keep]
+            if add_special_tokens:
+                ids = [0] + ids + [1]  # [CLS] ... [SEP]
+            input_ids = MockTensor([ids]) if return_tensors == "pt" else ids
             return {
-                "input_ids": MockTensor([[1]]),
+                "input_ids": input_ids,
-                "attention_mask": MockTensor([[1]]),
+                "attention_mask": MockTensor([[1] * len(ids)]),
             }
 
+        def num_special_tokens_to_add(self) -> int:
+            return 2
+
+        def decode(self, ids, skip_special_tokens=True):
+            kept = [i for i in ids if not (skip_special_tokens and i in (0, 1))]
+            return " ".join(str(i) for i in kept)
+
     # ── Mock torch ─────────────────────────────────────────────────
 
+    def _flat_values(x):
+        """Read a 1-D value list from a MockTensor of either orientation.
+
+        torch.tensor(flat_list) makes a (1, n) ROW, while softmax/argmax
+        results elsewhere are (n, 1) COLUMNS; the old readers only handled
+        the column shape, so a multi-category row silently collapsed to
+        its first value and every classify returned (category_0, 1.0).
+        """
+        if isinstance(x, MockTensor):
+            if len(x._data) == 1:
+                return list(x._data[0])
+            return [row[0] for row in x._data]
+        return list(x)
+
     def _softmax(x, dim=0):
         """Compute softmax over dim 0 using MockTensor values."""
-        if isinstance(x, MockTensor):
+        vals = _flat_values(x)
-            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)
@@ -297,12 +332,8 @@ def _install_fake_modules():
         return MockTensor([[p] for p in probs])
 
     def _argmax(x, dim=None):
-        if isinstance(x, MockTensor):
+        vals = _flat_values(x)
-            vals = [row[0] for row in x._data] if x._data else []
+        return MockTensor([vals.index(max(vals))])
-        else:
-            vals = list(x)
-        idx = vals.index(max(vals))
-        return MockTensor([idx])
 
     def _make_tensor(data):
         if isinstance(data, MockTensor):
@@ -344,9 +375,11 @@ def _clear_caches():
     """All caches in local_encoder must be empty between tests."""
     local_encoder._model_cache.clear()
     local_encoder._desc_cache.clear()
+    local_encoder._budget_fallback_logged = False
     yield
     local_encoder._model_cache.clear()
     local_encoder._desc_cache.clear()
+    local_encoder._budget_fallback_logged = False
 
 
 @pytest.fixture
@@ -648,4 +681,529 @@ def test_desc_cache_extended_for_new_categories():
     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
+    assert "debugging" in cached_after
+
+
+# ── Token-budget fitting (the 2026-09-18 truncation regression) ────────
+#
+# classify_zero_shot's final tokenizer call used truncation=True without
+# max_length, so HuggingFace clipped the prefixed task to the tokenizer's
+# model_max_length from the RIGHT: for BGE (BERT family, 512) a 6000+
+# char agent-session prompt became ~510 tokens of its shared head
+# boilerplate, the task-specific tail never reached the model, and every
+# long request returned the same verdict. Production evidence: six
+# distinct requests (6425-6572 chars) all classified diff_checking at
+# confidence 0.192363 for 20+ minutes.
+#
+# The fakes below emulate the HF contract faithfully enough that the
+# regression test FAILS against the pre-fix code and PASSES after.
+
+
+class SessionStyleTokenizer:
+    """Fake HF tokenizer with the contract that broke production.
+
+    - known ``model_max_length`` (512, the BERT/BGE window);
+    - right-side truncation when called with ``truncation=True`` and no
+      ``max_length`` (HF's default ``truncation_side='right'`` keeps the
+      head and silently drops the tail);
+    - plain Python id lists when called without ``return_tensors``;
+    - content-dependent ids: one id per distinct word, allocated in
+      first-seen order, with a ``decode`` that reconstructs text.
+    """
+
+    def __init__(self, model_max_length: int = 512) -> None:
+        self.model_max_length = model_max_length
+        self.truncation_side = "right"
+        self._vocab: dict = {}
+        self._words: dict = {}
+
+    @classmethod
+    def from_pretrained(cls, model_id, **kwargs):
+        return cls()
+
+    def _ids_for(self, text: str) -> list:
+        ids = []
+        for word in text.split():
+            if word not in self._vocab:
+                next_id = len(self._vocab) + 2  # 0=[CLS], 1=[SEP]
+                self._vocab[word] = next_id
+                self._words[next_id] = word
+            ids.append(self._vocab[word])
+        return ids
+
+    def __call__(
+        self, text, add_special_tokens=True, truncation=False,
+        padding=True, return_tensors=None, max_length=None, **_kw,
+    ):
+        ids = self._ids_for(str(text))
+        if truncation:
+            keep = self.model_max_length - (2 if add_special_tokens else 0)
+            ids = ids[:keep]
+        if add_special_tokens:
+            ids = [0] + ids + [1]
+        input_ids = MockTensor([ids]) if return_tensors == "pt" else ids
+        return {
+            "input_ids": input_ids,
+            "attention_mask": MockTensor([[1] * len(ids)]),
+        }
+
+    def num_special_tokens_to_add(self) -> int:
+        return 2
+
+    def decode(self, ids, skip_special_tokens=True):
+        words = [
+            self._words[i]
+            for i in ids
+            if not (skip_special_tokens and i in (0, 1))
+        ]
+        return " ".join(words)
+
+
+class ContentSignModel:
+    """Embeds a token sequence as the SIGN of its last content token id.
+
+    Provably identical verdicts for inputs that truncate to the same
+    prefix (same last surviving token), and opposite sides of the
+    centroid pair for the regression test's two tails: with 100 alpha
+    words vs 101 beta words after the same boilerplate, the last content
+    token takes ids of opposite parity by construction. The seeded
+    [[1.0]] (coding_general) / [[-1.0]] (debugging) descriptions turn
+    that sign into a category flip.
+    """
+
+    @classmethod
+    def from_pretrained(cls, model_id, **kwargs):
+        return cls()
+
+    def to(self, device_str):
+        return self
+
+    def __call__(self, **kwargs):
+        ids = kwargs["input_ids"]
+        if isinstance(ids, MockTensor):
+            ids = ids._data[0]
+        content = [i for i in ids if i not in (0, 1)]
+        last = float(content[-1]) if content else 0.0
+        sign = 1.0 if last % 2 == 0 else -1.0
+        return types.SimpleNamespace(last_hidden_state=MockTensor([[sign]]))
+
+
+def _session_style_prompt(tail_words: list) -> str:
+    """An agent-session-shaped prompt: huge shared head, varying tail."""
+    boilerplate = " ".join(
+        f"session guideline {i} assume tools available" for i in range(400)
+    )
+    return f"{boilerplate} {' '.join(tail_words)}"
+
+
+def test_long_shared_prefix_tasks_do_not_collapse_to_identical_verdict():
+    """THE regression tripwire: two long prompts sharing thousands of
+    chars of identical head boilerplate but differing only in tail
+    content must NOT produce the same (category, confidence).
+
+    Pre-fix, both prefixed tasks right-truncated to the same 510 shared
+    boilerplate tokens -> same embedding -> same verdict, exactly the
+    production signature. This fails against pre-fix code by
+    construction: the tail words sit beyond token 510 and are never seen
+    by the model. Post-fix the tail survives the fit and the verdicts
+    separate.
+    """
+    tail_a = [f"alpha{i}" for i in range(100)]
+    tail_b = [f"beta{i}" for i in range(101)]
+    task_a = _session_style_prompt(tail_a)
+    task_b = _session_style_prompt(tail_b)
+    assert len(task_a) > 2000 and len(task_b) > 2000
+    assert task_a.startswith("session guideline 0") == task_b.startswith(
+        "session guideline 0"
+    )
+    assert task_a[: -len(" ".join(tail_a)) - 1] == (
+        task_b[: -len(" ".join(tail_b)) - 1]
+    )
+
+    sys.modules["transformers"].AutoTokenizer = SessionStyleTokenizer
+    sys.modules["transformers"].AutoModel = ContentSignModel
+    local_encoder._desc_cache[("m", "cpu")] = {
+        "coding_general": MockTensor([[1.0]]),
+        "debugging": MockTensor([[-1.0]]),
+    }
+
+    cat_a, conf_a = local_encoder.classify_zero_shot(
+        task_a, ["coding_general", "debugging"], model_id="m", device="cpu",
+    )
+    cat_b, conf_b = local_encoder.classify_zero_shot(
+        task_b, ["coding_general", "debugging"], model_id="m", device="cpu",
+    )
+    assert (cat_a, conf_a) != (cat_b, conf_b)
+    assert cat_a != cat_b
+
+
+def test_fit_task_noop_when_under_budget():
+    """A task within the usable budget passes through unchanged."""
+    tok = SessionStyleTokenizer(model_max_length=512)
+    short = "summarize this paragraph carefully"
+    assert local_encoder._fit_task_to_token_budget(short, tok) == short
+
+
+def test_fit_task_over_budget_keeps_head_and_tail_drops_middle():
+    """Over budget: the fitted text keeps a head slice and the tail,
+    drops the middle, splices the elision marker, and the re-tokenized
+    prefixed result fits the model window."""
+    tok = SessionStyleTokenizer(model_max_length=512)
+    task = (
+        " ".join(f"head{i}" for i in range(200))
+        + " " + " ".join(f"mid{i}" for i in range(2000))
+        + " " + " ".join(f"tail{i}" for i in range(200))
+    )
+    fitted = local_encoder._fit_task_to_token_budget(task, tok)
+    assert fitted != task
+    assert "head0" in fitted and "head121" in fitted
+    assert "tail199" in fitted and "mid1999" in fitted
+    assert "mid1000" not in fitted
+    assert "[... content elided ...]" in fitted
+    total = tok(local_encoder._BGE_QUERY_INSTRUCTION + fitted)
+    assert len(total["input_ids"]) <= tok.model_max_length
+
+
+def test_fit_budget_subtracts_prefix_and_special_tokens():
+    """The usable budget is model_max_length - prefix tokens - specials:
+    a task of exactly that many tokens is a no-op, one token more is
+    fitted, and the fitted result fits the real window."""
+    tok = SessionStyleTokenizer(model_max_length=512)
+    prefix_words = len(local_encoder._BGE_QUERY_INSTRUCTION.split())
+    budget = (
+        tok.model_max_length
+        - prefix_words
+        - tok.num_special_tokens_to_add()
+    )
+    exact = " ".join(f"w{i}" for i in range(budget))
+    assert local_encoder._fit_task_to_token_budget(exact, tok) == exact
+
+    over = exact + " overflow"
+    fitted = local_encoder._fit_task_to_token_budget(over, tok)
+    assert fitted != over
+    total = tok(local_encoder._BGE_QUERY_INSTRUCTION + fitted)
+    assert len(total["input_ids"]) <= tok.model_max_length
+
+
+@pytest.mark.parametrize("bad_max_length", [None, int(1e30), 0, -5])
+def test_fit_falls_back_when_model_max_length_unusable(bad_max_length):
+    """A missing/sentinel/non-positive model_max_length falls back to the
+    conservative 512 window (logged once via the module flag) instead of
+    blowing up, and the fitted result still fits 512."""
+    tok = SessionStyleTokenizer(model_max_length=512)
+    tok.model_max_length = bad_max_length
+    task = " ".join(f"w{i}" for i in range(4000))
+    fitted = local_encoder._fit_task_to_token_budget(task, tok)
+    assert fitted != task
+    assert local_encoder._budget_fallback_logged
+    total = tok(local_encoder._BGE_QUERY_INSTRUCTION + fitted)
+    assert len(total["input_ids"]) <= local_encoder._FALLBACK_MODEL_MAX_LENGTH
+
+
+def test_fit_degenerate_window_returns_task_unchanged():
+    """A window smaller than prefix + specials leaves the task alone
+    rather than producing a nonsensical slice."""
+    tok = SessionStyleTokenizer(model_max_length=4)  # < prefix(7) + 2
+    task = "a b c"
+    assert local_encoder._fit_task_to_token_budget(task, tok) == task
+
+
+def test_classify_zero_shot_leaves_cached_tokenizer_state_untouched():
+    """The fit must not mutate the shared cached tokenizer (no
+    truncation_side flips) — the cache is process-wide."""
+    sys.modules["transformers"].AutoTokenizer = SessionStyleTokenizer
+    sys.modules["transformers"].AutoModel = ContentSignModel
+    local_encoder.classify_zero_shot(
+        _session_style_prompt([f"alpha{i}" for i in range(100)]),
+        ["coding_general", "debugging"],
+        model_id="m", device="cpu",
+    )
+    tok = local_encoder._model_cache[("m", "cpu")][1]
+    assert tok.truncation_side == "right"
+    assert tok.model_max_length == 512
+
+
+# ── Task-text isolation (the 2026-09-19 noise-dominance regression) ────
+#
+# classify_zero_shot embedded the RAW task: a prose instruction wrapped
+# in agent-session noise (fenced code block, "Tool result:" line,
+# <system-reminder> tag) mean-pooled the noise tokens with equal weight,
+# and the verdict scored the noise — measured on the real
+# BAAI/bge-large-en-v1.5, 8 clean tasks scored 8/8 but the same
+# instructions wrapped scored 2/8, on SHORT inputs far under any
+# truncation limit (a second, distinct bug; the fit fix does not touch
+# it — windowing alone measured 2/8 on long noisy pairs too).
+#
+# The fakes below reproduce that mechanism in miniature: tokens drawn
+# from the code/structural vocabulary push the pooled embedding toward
+# the code centroid, prose tokens toward the prose centroid, so
+# noise-count dominance flips the verdict exactly like the real model.
+
+_INSTRUCTION = (
+    "Please write a clear docstring and a short readme section "
+    "explaining the storage module."
+)
+
+_NOISE_REMINDER = (
+    "<system-reminder>\nsession context: agent harness active\n"
+    "</system-reminder>"
+)
+_NOISE_CODE = (
+    "```python\ndef parse_line(line):\n"
+    "    tokens = line.strip().split(',')\n"
+    "    if not tokens:\n"
+    "        return []\n"
+    "    return [int(t) for t in tokens if t.isdigit()]\n"
+    "FILES = {'a': 1, 'b': 2}\n```"
+)
+_NOISE_TOOL = "Tool result: file read successfully, 12 lines"
+
+
+def _norm(word: str) -> str:
+    """Normalize a whitespace token to its word characters (lowercased)."""
+    return "".join(ch for ch in word.lower() if ch.isalnum() or ch == "_")
+
+
+# Every multi-char word of the noise fixture is a "code/structural"
+# vocabulary entry — the fake tokenizer's stand-in for "this token
+# embeds near code", mirroring how real code/tool tokens pull the
+# pooled embedding toward code-flavored categories.
+_CODE_VOCAB = frozenset(
+    w
+    for w in (
+        _norm(tok)
+        for tok in (_NOISE_REMINDER + " " + _NOISE_CODE + " " + _NOISE_TOOL).split()
+    )
+    if len(w) > 1
+)
+
+
+class NoiseDominantTokenizer:
+    """Fake HF tokenizer whose ids encode word class.
+
+    Words in _CODE_VOCAB get ids >= 1000; prose words get ids from 2 up
+    (first-seen, like SessionStyleTokenizer). Otherwise the same HF
+    contract: known model_max_length, right-side truncation under
+    ``truncation=True`` without ``max_length``, plain id lists without
+    ``return_tensors``, and a decode that reconstructs text.
+    """
+
+    def __init__(self, model_max_length: int = 512) -> None:
+        self.model_max_length = model_max_length
+        self.truncation_side = "right"
+        self._vocab: dict = {}
+        self._words: dict = {}
+
+    @classmethod
+    def from_pretrained(cls, model_id, **kwargs):
+        return cls()
+
+    def _ids_for(self, text: str) -> list:
+        ids = []
+        for word in str(text).split():
+            norm = _norm(word)
+            if not norm:
+                continue
+            if norm not in self._vocab:
+                if norm in _CODE_VOCAB:
+                    next_id = 1000 + len(self._vocab)
+                else:
+                    next_id = 2 + len(self._vocab)
+                self._vocab[norm] = next_id
+                self._words[next_id] = word
+            ids.append(self._vocab[norm])
+        return ids
+
+    def __call__(
+        self, text, add_special_tokens=True, truncation=False,
+        padding=True, return_tensors=None, max_length=None, **_kw,
+    ):
+        ids = self._ids_for(text)
+        if truncation:
+            keep = self.model_max_length - (2 if add_special_tokens else 0)
+            ids = ids[:keep]
+        if add_special_tokens:
+            ids = [0] + ids + [1]
+        input_ids = MockTensor([ids]) if return_tensors == "pt" else ids
+        return {
+            "input_ids": input_ids,
+            "attention_mask": MockTensor([[1] * len(ids)]),
+        }
+
+    def num_special_tokens_to_add(self) -> int:
+        return 2
+
+    def decode(self, ids, skip_special_tokens=True):
+        words = [
+            self._words[i]
+            for i in ids
+            if not (skip_special_tokens and i in (0, 1))
+        ]
+        return " ".join(words)
+
+
+class NoiseDominantModel:
+    """Embeds a token sequence as the MEAN of per-token class values.
+
+    +1.0 for code-vocabulary ids (>= 1000), -1.0 for prose ids: mean
+    pooling weights every token equally, so whichever class supplies
+    more tokens wins the pooled sign — the miniature of the real
+    failure, where code/tool noise outnumbers the instruction words and
+    the verdict follows the noise.
+    """
+
+    @classmethod
+    def from_pretrained(cls, model_id, **kwargs):
+        return cls()
+
+    def to(self, device_str):
+        return self
+
+    def __call__(self, **kwargs):
+        ids = kwargs["input_ids"]
+        if isinstance(ids, MockTensor):
+            ids = ids._data[0]
+        content = [i for i in ids if i not in (0, 1)]
+        values = [1.0 if i >= 1000 else -1.0 for i in content]
+        mean = sum(values) / len(values) if values else 0.0
+        return types.SimpleNamespace(last_hidden_state=MockTensor([[mean]]))
+
+
+def _noisy_instruction() -> str:
+    """The clean instruction wrapped in the operator's repro noise:
+    system-reminder tag + fenced code block + tool-result line, the
+    instruction still clearly present at the END."""
+    return "\n\n".join([_NOISE_REMINDER, _NOISE_CODE, _NOISE_TOOL, _INSTRUCTION])
+
+
+def test_noise_wrapped_task_classifies_same_as_clean_pair():
+    """THE noise tripwire: the same prose instruction, bare and wrapped
+    in agent-session noise, must classify to the SAME category.
+
+    Fails against pre-isolation code by construction: the wrapped task's
+    code/structural tokens outnumber the prose tokens, so the pooled
+    mean goes positive (code centroid) while the bare instruction pools
+    negative (prose centroid) — a category flip, the miniature of the
+    real-model observation. Post-isolation the noise shapes are stripped
+    before embedding and both tasks pool to the prose centroid.
+    """
+    clean = _INSTRUCTION
+    noisy = _noisy_instruction()
+    # Precondition: the fixture is short — windowing cannot apply, so
+    # this pair exercises isolation, not the fit.
+    assert len(noisy) < 1000
+    assert not (_CODE_VOCAB & {_norm(w) for w in clean.split()})
+
+    sys.modules["transformers"].AutoTokenizer = NoiseDominantTokenizer
+    sys.modules["transformers"].AutoModel = NoiseDominantModel
+    local_encoder._desc_cache[("m", "cpu")] = {
+        "docs_writing": MockTensor([[-1.0]]),
+        "coding_general": MockTensor([[1.0]]),
+    }
+
+    cat_clean, _ = local_encoder.classify_zero_shot(
+        clean, ["docs_writing", "coding_general"], model_id="m", device="cpu",
+    )
+    cat_noisy, _ = local_encoder.classify_zero_shot(
+        noisy, ["docs_writing", "coding_general"], model_id="m", device="cpu",
+    )
+    assert cat_clean == "docs_writing"
+    assert cat_noisy == cat_clean, (
+        f"noise flipped the verdict: clean={cat_clean}, noisy={cat_noisy}"
+    )
+
+
+def test_isolate_strips_fenced_code_blocks():
+    """A fenced code block is removed outright — body AND fences."""
+    task = "Summarize the setup steps.\n```python\nimport os\nos.getcwd()\n```"
+    out = local_encoder._isolate_task_text(task)
+    assert "import os" not in out
+    assert "os.getcwd()" not in out
+    assert "```" not in out
+
+
+def test_isolate_removal_decision_no_placeholder():
+    """The eval-decided placeholder-vs-removal winner is REMOVAL: the
+    stripped text must not carry any '[code]'-style placeholder — the
+    placeholder token itself pulled long noisy pairs to 6/8 vs removal's
+    8/8 on the real model."""
+    task = (
+        "Explain what this algorithm computes and when you would use it.\n"
+        "```\nfor i in range(3):\n    print(i)\n```"
+    )
+    out = local_encoder._isolate_task_text(task)
+    assert "[code]" not in out
+    assert "[elided]" not in out
+    assert "for i in range(3)" not in out
+
+
+def test_isolate_strips_tool_result_lines():
+    """'Tool result:'-shaped lines are stripped line-level; result and
+    output/call variants share the marker shape."""
+    task = (
+        "Tool result: file read successfully, 12 lines\n"
+        "Tool output: exit code 0\n"
+        "Now summarize the findings above."
+    )
+    out = local_encoder._isolate_task_text(task)
+    assert "Tool result:" not in out
+    assert "Tool output:" not in out
+    assert "exit code 0" not in out
+    assert "summarize the findings" in out
+
+
+def test_isolate_strips_system_reminder_content():
+    """<system-reminder> spans are removed with their content; the tag
+    itself does not survive either."""
+    task = (
+        "<system-reminder>session id 42, tools enabled</system-reminder>\n"
+        "Translate the attached paragraph to French."
+    )
+    out = local_encoder._isolate_task_text(task)
+    assert "session id 42" not in out
+    assert "system-reminder" not in out
+    assert "Translate the attached paragraph" in out
+
+
+def test_isolate_preserves_trailing_instruction():
+    """The prose instruction survives verbatim when it trails the
+    noise — the operator's repro shape."""
+    assert _noisy_instruction().endswith(_INSTRUCTION)
+    out = local_encoder._isolate_task_text(_noisy_instruction())
+    assert out.strip().endswith(_INSTRUCTION)
+
+
+def test_isolate_keeps_inline_code_spans():
+    """Inline code spans are KEPT by default: instructions reference
+    real file names and commands, which are signal, not noise."""
+    task = "Rename `config_loader.py` to `settings.py` and fix the imports."
+    assert local_encoder._isolate_task_text(task) == task
+
+
+def test_isolate_floor_guard_falls_back_on_all_code_task():
+    """A task that is essentially only a code dump (nothing left above
+    the character floor) falls back to the original text rather than
+    embedding a stripped stub."""
+    bare_dump = "```python\n" + "\n".join(f"x{i} = {i}" for i in range(40)) + "\n```"
+    assert local_encoder._isolate_task_text(bare_dump) == bare_dump
+
+
+def test_isolate_is_idempotent():
+    """Applying isolation twice equals applying it once."""
+    once = local_encoder._isolate_task_text(_noisy_instruction())
+    assert local_encoder._isolate_task_text(once) == once
+
+
+def test_isolate_unterminated_fence_keeps_leading_instruction():
+    """An unterminated fence (truncated tool output) eats to the end of
+    the string, so a leading instruction survives and a trailing one
+    cannot — the floor guard then reverts rather than embed a stub."""
+    leading = (
+        "Summarize the parser module.\n\n```python\nimport csv\nrows = []\n"
+    )
+    out = local_encoder._isolate_task_text(leading)
+    assert "Summarize the parser module." in out
+    assert "import csv" not in out
+
+    trailing = "```python\nimport csv\n" + "x = 1\n" * 30 + "summarize it"
+    assert local_encoder._isolate_task_text(trailing) == trailing