feat(profiler/behave_shell): emit cognitive.inter_command_consistency

BEHAVE-EXTRACTOR.md Phase A Step 8. Dispersion / bimodality of
inter-command pauses. HUMAN-bimodal vs LLM-metronomic.

* _features/cognitive.py:inter_command_consistency(ctx) emits one
  Observation in {metronomic, variable, bimodal}.
* CV = stdev / mean of ctx.inter_cmd_iats. CV < 0.40 → metronomic
  (LLM-pure; corpus anchor 0.24); CV ≥ 1.50 → bimodal heuristic
  (LLM-assisted human; v0.1 placeholder, true bimodal via Hartigan
  dip is registry-flagged for v0.2); else → variable (human;
  corpus anchor 0.94).
* < 2 IATs or zero mean → skip emission. < 5 commands halves
  confidence (0.40 vs 0.75) per sample-size honesty.

Tests: too-few IATs → no emission, uniform → metronomic,
human-like dispersion → variable, extreme bursts+gaps → bimodal,
low-sample-count → reduced confidence.

Step 8 closes the six-primitive calibration floor for Phase A.
Step 9 (calibration grid lockdown) is the gate that pins it.

decnet/profiler/behave_shell/_features/__init__.py

@@ -14,6 +14,7 @@ from decnet.profiler.behave_shell._ctx import SessionContext
 from decnet.profiler.behave_shell._features.cognitive import (
     command_branch_diversity,
     feedback_loop_engagement,
+    inter_command_consistency,
     inter_command_latency_class,
 )
 from decnet.profiler.behave_shell._features.motor import (
@@ -29,4 +30,5 @@ FEATURES: tuple[FeatureFn, ...] = (
     inter_command_latency_class,
     command_branch_diversity,
     feedback_loop_engagement,
+    inter_command_consistency,
 )

decnet/profiler/behave_shell/_features/cognitive.py

@@ -24,6 +24,8 @@ from decnet.profiler.behave_shell._thresholds import (
     INTER_CMD_LLM_LIGHTWEIGHT_MAX,
     INTER_CMD_TYPING_MAX,
     MIN_COMMANDS_FOR_FULL_CONFIDENCE,
+    PAUSE_CV_BIMODAL_MIN,
+    PAUSE_CV_METRONOMIC_MAX,
 )
 
 
@@ -152,3 +154,40 @@ def feedback_loop_engagement(ctx: SessionContext) -> Iterator[Observation]:
         value=value,
         confidence=0.75,
     )
+
+
+def inter_command_consistency(ctx: SessionContext) -> Iterator[Observation]:
+    """Emit ``cognitive.inter_command_consistency``.
+
+    CV (stdev / mean) of inter-command IATs.
+
+    * ``metronomic`` (CV < 0.40) → LLM-pure. Empirical anchor:
+      LLM-simulated session CV ≈ 0.24 in this corpus.
+    * ``variable`` (0.40 ≤ CV < 1.50) → human. Empirical anchor:
+      human session CV ≈ 0.94.
+    * ``bimodal`` (CV ≥ 1.50) → LLM-assisted human, heuristic. v0.1
+      uses CV-only; true bimodal detection (Hartigan dip / two-peak)
+      is filed for v0.2 per the registry's ``notes:`` field.
+    """
+    iats = ctx.inter_cmd_iats
+    if len(iats) < 2:
+        return
+    mean = statistics.fmean(iats)
+    if mean <= 0.0:
+        return
+    cv = statistics.stdev(iats) / mean
+    if cv < PAUSE_CV_METRONOMIC_MAX:
+        value = "metronomic"
+    elif cv >= PAUSE_CV_BIMODAL_MIN:
+        value = "bimodal"
+    else:
+        value = "variable"
+    confidence = (
+        0.40 if len(ctx.commands) < MIN_COMMANDS_FOR_FULL_CONFIDENCE else 0.75
+    )
+    yield make_observation(
+        ctx,
+        primitive="cognitive.inter_command_consistency",
+        value=value,
+        confidence=confidence,
+    )

tests/profiler/behave_shell/test_cognitive_inter_command_consistency.py

@@ -0,0 +1,57 @@
+"""Step 8: ``cognitive.inter_command_consistency``."""
+from __future__ import annotations
+
+from decnet.profiler.behave_shell import extract_session
+from decnet.profiler.behave_shell._parse import AsciinemaEvent
+
+
+def _of(observations: list, primitive: str):
+    obs = [o for o in observations if o.primitive == primitive]
+    assert len(obs) == 1, f"expected exactly one {primitive}, got {len(obs)}"
+    return obs[0]
+
+
+def _commands_at(starts: list[float]) -> list[AsciinemaEvent]:
+    events: list[AsciinemaEvent] = []
+    for s in starts:
+        events.append((s, "i", "x\r"))
+    return events
+
+
+def test_too_few_iats_no_emission() -> None:
+    out = list(extract_session(_commands_at([0.0, 1.0]), sid="cv-low"))
+    assert [o for o in out if o.primitive == "cognitive.inter_command_consistency"] == []
+
+
+def test_uniform_pace_emits_metronomic() -> None:
+    # Constant 1s gap → CV 0
+    out = list(extract_session(
+        _commands_at([i * 1.0 for i in range(8)]), sid="cv-metro",
+    ))
+    obs = _of(out, "cognitive.inter_command_consistency")
+    assert obs.value == "metronomic"
+
+
+def test_human_like_dispersion_emits_variable() -> None:
+    # Pauses around 1s mean with CV ≈ 0.9 (human empirical)
+    starts = [0.0, 0.4, 1.4, 1.6, 4.0, 4.4, 7.5]
+    out = list(extract_session(_commands_at(starts), sid="cv-var"))
+    obs = _of(out, "cognitive.inter_command_consistency")
+    assert obs.value == "variable"
+
+
+def test_extreme_dispersion_emits_bimodal() -> None:
+    # Mix of very tight bursts and very long gaps → CV well above 1.5
+    starts = [0.0, 0.1, 0.2, 30.0, 30.1, 30.2, 60.0]
+    out = list(extract_session(_commands_at(starts), sid="cv-bi"))
+    obs = _of(out, "cognitive.inter_command_consistency")
+    assert obs.value == "bimodal"
+
+
+def test_low_sample_count_reduces_confidence() -> None:
+    # 3 commands → 2 IATs; below the floor of 5
+    short = list(extract_session(_commands_at([0.0, 1.0, 2.0]), sid="cv-short"))
+    full = list(extract_session(_commands_at([i * 1.0 for i in range(8)]), sid="cv-full"))
+    s = _of(short, "cognitive.inter_command_consistency")
+    f = _of(full, "cognitive.inter_command_consistency")
+    assert s.confidence < f.confidence