fix(skill-versioning): consume divergence_tolerance, replace vacuous catalog-isolation guard

- divergence_tolerance was loaded from config but diff_flow_versions
  never consulted it, so structural_divergence was always plain
  sequence-equality regardless of the configured tolerance. Now
  actually applied per design.md D4.
- The skill-catalog-subscription isolation guard test asserted
  'skills.catalog' not in sys.modules, which is vacuously true since
  that module doesn't exist anywhere yet. Replaced with a real check
  against store.py's actual imports.

openspec: skill-versioning capability, archived change skill-learning-runtime

skills_learning/versioning.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from dataclasses import dataclass
 from typing import Any
 
+from skills_learning.config import SkillAuthoringConfig
 from skills_learning.models import FlowStep, FlowTemplateSkill
 from skills_learning.store import SkillStore
 
@@ -13,6 +14,68 @@ class VersionDiff:
     stored_sequence: list[str]
     executed_sequence: list[str]
     argument_differences: list[tuple[int, str, Any, Any]]
+    argument_divergence_fraction: float
+
+
+def diff_flow_versions(
+    stored_steps: list[FlowStep],
+    executed_steps: list[FlowStep],
+    *,
+    config: SkillAuthoringConfig | None = None,
+) -> VersionDiff:
+    """Compare a stored skill's steps against a newly executed run.
+
+    Per design.md D4: a change to the tool-name *sequence* itself
+    (insertion/deletion/reorder) always triggers a new version, with no
+    tolerance applied. When the sequence matches (same skeleton), the
+    fraction of step positions whose arguments differ is compared against
+    ``config.divergence_tolerance`` (a fraction in ``[0, 1]``): a
+    "materially different parameter set" that exceeds that tolerance is
+    also treated as divergence worth a new version, per proposal.md's
+    "differ beyond a configured tolerance ... or a materially different
+    parameter set" trigger. When no ``config`` is supplied, tolerance is
+    treated as unlimited (matching the pre-existing, backward-compatible
+    behavior of never bumping a version for argument-only differences).
+    """
+    stored_sequence = [step.tool_name for step in stored_steps]
+    executed_sequence = [step.tool_name for step in executed_steps]
+    sequence_diverged = stored_sequence != executed_sequence
+
+    differences: list[tuple[int, str, Any, Any]] = []
+    diverged_positions = 0
+    total_positions = 0
+    if not sequence_diverged:
+        for index, (stored_step, executed_step) in enumerate(
+            zip(stored_steps, executed_steps, strict=True)
+        ):
+            total_positions += 1
+            keys = set(stored_step.args) | set(executed_step.args)
+            position_diverged = False
+            for key in sorted(keys):
+                stored_value = stored_step.args.get(key)
+                executed_value = executed_step.args.get(key)
+                if stored_value != executed_value:
+                    differences.append((index, key, stored_value, executed_value))
+                    position_diverged = True
+            if position_diverged:
+                diverged_positions += 1
+
+    argument_divergence_fraction = (
+        diverged_positions / total_positions if total_positions else 0.0
+    )
+    beyond_tolerance = (
+        not sequence_diverged
+        and config is not None
+        and argument_divergence_fraction > config.divergence_tolerance
+    )
+
+    return VersionDiff(
+        structural_divergence=sequence_diverged or beyond_tolerance,
+        stored_sequence=stored_sequence,
+        executed_sequence=executed_sequence,
+        argument_differences=differences,
+        argument_divergence_fraction=argument_divergence_fraction,
+    )
 
 
 @dataclass(frozen=True)
@@ -21,41 +84,17 @@ class VersioningResult:
     created_new_version: bool
 
 
-def diff_flow_versions(
-    stored_steps: list[FlowStep],
-    executed_steps: list[FlowStep],
-) -> VersionDiff:
-    stored_sequence = [step.tool_name for step in stored_steps]
-    executed_sequence = [step.tool_name for step in executed_steps]
-    structural = stored_sequence != executed_sequence
-    differences: list[tuple[int, str, Any, Any]] = []
-    if not structural:
-        for index, (stored_step, executed_step) in enumerate(
-            zip(stored_steps, executed_steps, strict=True)
-        ):
-            keys = set(stored_step.args) | set(executed_step.args)
-            for key in sorted(keys):
-                stored_value = stored_step.args.get(key)
-                executed_value = executed_step.args.get(key)
-                if stored_value != executed_value:
-                    differences.append((index, key, stored_value, executed_value))
-    return VersionDiff(
-        structural_divergence=structural,
-        stored_sequence=stored_sequence,
-        executed_sequence=executed_sequence,
-        argument_differences=differences,
-    )
-
-
 def store_synthesized_skill(
     store: SkillStore,
     candidate: FlowTemplateSkill,
+    *,
+    config: SkillAuthoringConfig | None = None,
 ) -> VersioningResult:
     latest = store.get_latest_by_name(candidate.name)
     if latest is None:
         return VersioningResult(store.create_version(candidate), True)
 
-    diff = diff_flow_versions(latest.steps, candidate.steps)
+    diff = diff_flow_versions(latest.steps, candidate.steps, config=config)
     if diff.structural_divergence:
         return VersioningResult(store.create_version(candidate, parent=latest), True)