fix(cognition): bound speculative-subject cache + evict on COHERENT promotion (#85)

Closes audit Finding 5 (2026-05-20).

Pre-fix ``CognitiveTurnPipeline._speculative_subjects`` was a bare
``set[str]`` that only grew over a session.  Two correctness gaps:

  * A subject promoted to ``EpistemicStatus.COHERENT`` via the teaching
    review loop kept appearing with the "(speculative, not yet
    reviewed)" marker forever, contaminating reviewed material on
    later probes.
  * Long teaching sessions widened the per-turn substring scan in
    ``_should_mark_speculative`` without bound.

Fix:

  * Back the cache with ``OrderedDict[str, None]`` (LRU) capped at
    ``_MAX_SPECULATIVE_SUBJECTS = 64``.
  * Introduce ``_remember_speculative_subject`` (insert / refresh) and
    ``_forget_speculative_subject`` (evict) helpers; route all
    SPECULATIVE inserts through them.
  * When a proposal lands as ``EpistemicStatus.COHERENT``, evict the
    subject and every long-enough non-stopword token derived from it,
    so the marker stops appearing on reviewed material.

Iteration order in ``_should_mark_speculative`` is unchanged (keys
view); lookups remain O(1).  No surface change for any case the prior
behavior didn't already mishandle, so byte-identical eval surfaces
stay stable (verified locally against ``core eval cognition`` public /
holdout / dev splits — all unchanged from MEMORY baseline).

Tests (7 new, ``tests/test_speculative_subject_lifecycle.py``):

  * storage is an OrderedDict and the cap is 64
  * remember normalizes (lower+strip) and drops empty input
  * remember refreshes LRU position on re-insert
  * cache caps at 64 with insertion-order eviction
  * forget is case-insensitive and removes the entry
  * forget on a missing / empty subject is a no-op
  * ``_should_mark_speculative`` triggers after remember and stops
    triggering after forget

Audit findings referenced:
https://github.com/AssetOverflow/core/pull/76 (Finding 5, "Unbounded
``_speculative_subjects``")

core/cognition/pipeline.py

@@ -15,6 +15,8 @@ Constraint: ChatRuntime.chat() and ChatResponse contract are unchanged.
 
 from __future__ import annotations
 
+from collections import OrderedDict
+
 from field.state import FieldState
 from core.cognition.result import CognitiveTurnResult
 from core.cognition.surface_resolution import resolve_surface
@@ -75,6 +77,15 @@ _SUBJECT_STOPWORDS: frozenset[str] = frozenset({
     "your", "their", "answer",
 })
 
+# Finding 5 (audit 2026-05-20) — cap the speculative-subjects cache so a
+# long teaching session cannot grow it without bound.  64 is large enough
+# to cover every distinct teaching subject a single session realistically
+# emits and small enough that the per-turn substring scan in
+# ``_should_mark_speculative`` stays trivially cheap.  LRU eviction: a
+# subject re-encountered as SPECULATIVE refreshes its position; coherent
+# promotion removes it explicitly.
+_MAX_SPECULATIVE_SUBJECTS = 64
+
 
 class CognitiveTurnPipeline:
     """Thin pipeline wrapper over ChatRuntime.
@@ -94,7 +105,16 @@ class CognitiveTurnPipeline:
         # (by subject substring or reflexive query shape), the surface
         # is prefixed with _SPECULATIVE_SURFACE_MARKER so the user can
         # tell ratified knowledge from unreviewed teaching material.
-        self._speculative_subjects: set[str] = set()
+        #
+        # Finding 5 (audit 2026-05-20) — backed by an OrderedDict so the
+        # cache is bounded (LRU, cap ``_MAX_SPECULATIVE_SUBJECTS``) and
+        # supports explicit eviction when a proposal is promoted to
+        # COHERENT.  Pre-fix this was a bare ``set`` that only grew,
+        # which both leaked speculative markers onto reviewed subjects
+        # forever and widened the per-turn substring scan unboundedly.
+        # Iteration order matches insertion / refresh order; lookups
+        # remain O(1).
+        self._speculative_subjects: OrderedDict[str, None] = OrderedDict()
 
     # ------------------------------------------------------------------
     # Public API
@@ -204,19 +224,28 @@ class CognitiveTurnPipeline:
         # proposal.subject (often a fragment of the correction text);
         # also split-and-add each ≥4-char token so prefixed parses like
         # "correction: wisdom" still match a probe about "wisdom".
-        if proposal is not None and proposal.epistemic_status is EpistemicStatus.SPECULATIVE:
+        if proposal is not None:
             sources: list[str] = []
             if proposal.triple is not None and proposal.triple[0]:
                 sources.append(proposal.triple[0])
             if proposal.subject:
                 sources.append(proposal.subject)
-            for src in sources:
-                lowered = src.lower().strip()
-                if lowered:
-                    self._speculative_subjects.add(lowered)
-                for tok in _SUBJECT_SPLIT_RE.split(src.lower()):
-                    if len(tok) >= 4 and tok not in _SUBJECT_STOPWORDS:
-                        self._speculative_subjects.add(tok)
+            if proposal.epistemic_status is EpistemicStatus.SPECULATIVE:
+                for src in sources:
+                    self._remember_speculative_subject(src)
+                    for tok in _SUBJECT_SPLIT_RE.split(src.lower()):
+                        if len(tok) >= 4 and tok not in _SUBJECT_STOPWORDS:
+                            self._remember_speculative_subject(tok)
+            elif proposal.epistemic_status is EpistemicStatus.COHERENT:
+                # Finding 5 (audit 2026-05-20) — once teaching review
+                # promotes a proposal to COHERENT, the subject is no
+                # longer speculative; evict its tokens so the marker
+                # stops appearing on subsequent probes about it.
+                for src in sources:
+                    self._forget_speculative_subject(src)
+                    for tok in _SUBJECT_SPLIT_RE.split(src.lower()):
+                        if len(tok) >= 4 and tok not in _SUBJECT_STOPWORDS:
+                            self._forget_speculative_subject(tok)
 
         # Advance turn counter and remember surface for next correction binding
         self._turn_number += 1
@@ -348,6 +377,34 @@ class CognitiveTurnPipeline:
             )
         return ratify_intent(intent, prompt_versor, vocab=vocab)
 
+    def _remember_speculative_subject(self, subject: str) -> None:
+        """Add (or refresh LRU position of) a speculative subject token.
+
+        Finding 5 (audit 2026-05-20).  Caps the cache at
+        ``_MAX_SPECULATIVE_SUBJECTS`` via insertion-order eviction.
+        Empty / whitespace-only inputs are dropped silently so callers
+        can pass raw fragments without guarding.
+        """
+        subject = subject.lower().strip()
+        if not subject:
+            return
+        self._speculative_subjects.pop(subject, None)
+        self._speculative_subjects[subject] = None
+        while len(self._speculative_subjects) > _MAX_SPECULATIVE_SUBJECTS:
+            self._speculative_subjects.popitem(last=False)
+
+    def _forget_speculative_subject(self, subject: str) -> None:
+        """Evict a subject from the speculative-marker cache.
+
+        Called when a SPECULATIVE proposal is promoted to COHERENT via
+        the teaching review loop, so reviewed material stops being
+        marked speculative on later probes.  No-op if the subject is
+        not present.
+        """
+        subject = subject.lower().strip()
+        if subject:
+            self._speculative_subjects.pop(subject, None)
+
     def _should_mark_speculative(self, text: str, surface: str) -> bool:
         """Decide whether ``surface`` should carry the SPECULATIVE marker.
 

tests/test_speculative_subject_lifecycle.py

@@ -0,0 +1,93 @@
+"""Speculative-subject cache lifecycle (Finding 5, audit 2026-05-20).
+
+Pre-fix ``CognitiveTurnPipeline._speculative_subjects`` was an unbounded
+``set[str]`` that only grew over a session.  Two correctness gaps:
+
+  * A subject promoted to ``EpistemicStatus.COHERENT`` via the teaching
+    review loop kept appearing with the "(speculative, not yet reviewed)"
+    marker forever.
+  * Long teaching sessions widened the per-turn substring scan inside
+    ``_should_mark_speculative`` without bound.
+
+These tests pin the lifecycle so the fix cannot silently regress.
+"""
+
+from __future__ import annotations
+
+from collections import OrderedDict
+
+import pytest
+
+from chat.runtime import ChatRuntime
+from core.cognition import CognitiveTurnPipeline
+from core.cognition.pipeline import _MAX_SPECULATIVE_SUBJECTS
+
+
+@pytest.fixture()
+def pipeline() -> CognitiveTurnPipeline:
+    return CognitiveTurnPipeline(runtime=ChatRuntime())
+
+
+def test_storage_is_bounded_ordereddict(pipeline: CognitiveTurnPipeline) -> None:
+    assert isinstance(pipeline._speculative_subjects, OrderedDict)
+    assert _MAX_SPECULATIVE_SUBJECTS == 64
+
+
+def test_remember_inserts_and_normalizes(pipeline: CognitiveTurnPipeline) -> None:
+    pipeline._remember_speculative_subject("  TRUTH  ")
+    assert "truth" in pipeline._speculative_subjects
+    pipeline._remember_speculative_subject("")
+    pipeline._remember_speculative_subject("   ")
+    # Empty / whitespace inputs are silently dropped
+    assert list(pipeline._speculative_subjects) == ["truth"]
+
+
+def test_remember_refreshes_lru_position(pipeline: CognitiveTurnPipeline) -> None:
+    pipeline._remember_speculative_subject("alpha")
+    pipeline._remember_speculative_subject("beta")
+    pipeline._remember_speculative_subject("alpha")
+    # alpha was re-touched after beta, so alpha is now the MRU entry
+    assert list(pipeline._speculative_subjects) == ["beta", "alpha"]
+
+
+def test_cache_caps_at_max(pipeline: CognitiveTurnPipeline) -> None:
+    for i in range(_MAX_SPECULATIVE_SUBJECTS + 10):
+        pipeline._remember_speculative_subject(f"subj{i:03d}")
+    assert len(pipeline._speculative_subjects) == _MAX_SPECULATIVE_SUBJECTS
+    # Oldest 10 entries (subj000..subj009) evicted; newest survive
+    keys = list(pipeline._speculative_subjects)
+    assert "subj000" not in keys
+    assert "subj009" not in keys
+    assert "subj010" in keys
+    assert keys[-1] == f"subj{_MAX_SPECULATIVE_SUBJECTS + 9:03d}"
+
+
+def test_forget_removes(pipeline: CognitiveTurnPipeline) -> None:
+    pipeline._remember_speculative_subject("truth")
+    pipeline._remember_speculative_subject("light")
+    pipeline._forget_speculative_subject("Truth")  # case-insensitive
+    assert "truth" not in pipeline._speculative_subjects
+    assert "light" in pipeline._speculative_subjects
+
+
+def test_forget_missing_is_noop(pipeline: CognitiveTurnPipeline) -> None:
+    pipeline._remember_speculative_subject("truth")
+    pipeline._forget_speculative_subject("never-added")
+    pipeline._forget_speculative_subject("")
+    assert list(pipeline._speculative_subjects) == ["truth"]
+
+
+def test_should_mark_speculative_still_iterates_correctly(
+    pipeline: CognitiveTurnPipeline,
+) -> None:
+    """The marker decision still triggers on stored subjects."""
+    pipeline._remember_speculative_subject("wisdom")
+    assert pipeline._should_mark_speculative(
+        text="Tell me about wisdom",
+        surface="Wisdom is the application of knowledge.",
+    )
+    pipeline._forget_speculative_subject("wisdom")
+    assert not pipeline._should_mark_speculative(
+        text="Tell me about wisdom",
+        surface="Wisdom is the application of knowledge.",
+    )