feat(persona): load the falsification skill from commons at call time

[skip-docs] — the invariant row for this plan lands in Step 13, after the mutations.

Amendment 2 in the same commit: the framework-neutrality sweep covered only
expert-reviewer, so the skill that arrived by subtree pull had no framework guard
anywhere. GUARDED_SKILL_DIRS names both, and a fail-closed coverage arm turns red
when a future pull brings a third skill that is not listed.

Co-Authored-By: Claude <claude-opus-5>

src/portfolio_optimiser/persona.py

@@ -54,3 +54,87 @@ def load_persona_example() -> PersonaExample:
         rationale=data["rationale"],
         marker=data["marker"],
     )
+
+
+# --- The falsification-reviewer skill (Step 12, D4) ----------------------------------------------
+# A SECOND shared Agent Skill, read the same way and for the same reason: it is authored in commons
+# and arrives by ``git subtree pull``, so a loader that consumes it is what keeps it from rotting
+# silently. Same call-time seam, same fail-fast discipline as the persona example above.
+
+_FALSIFICATION_EXAMPLE_PATH: Path | None = None
+
+_FALSIFICATION_SUBPATH = (
+    Path("skills") / "falsification-reviewer" / "references" / "example-evidence.json"
+)
+
+
+def _falsification_example_path() -> Path:
+    """The worked example's location: the test seam if set, else resolved under ``shared_root()``."""
+    if _FALSIFICATION_EXAMPLE_PATH is not None:
+        return _FALSIFICATION_EXAMPLE_PATH
+    return shared_root() / _FALSIFICATION_SUBPATH
+
+
+@dataclass(frozen=True)
+class FalsificationConcept:
+    """One concept the worked example judged, with what the readers should derive from its bytes.
+
+    ``frontmatter_verbatim`` is AUTHORITATIVE and is the only field a consumer should materialise
+    from. The example also ships a line-oriented ``frontmatter`` projection, which is informative
+    only: by construction it cannot carry a block form, so materialising from it would turn the
+    unreadable case into an absent one — the tolerant read that widens the answer.
+    """
+
+    concept_id: str
+    #: The document's frontmatter block exactly as written, newlines and indentation included.
+    frontmatter_verbatim: str
+    state: str
+    reason: str | None
+    items_seen: int
+    trust_tier: str
+    adjudication: str
+    #: Whether the verdict was allowed to rest on this concept.
+    relied_on: bool
+
+
+@dataclass(frozen=True)
+class FalsificationExample:
+    """The falsification reviewer's worked example: one claim, the concepts consulted, the verdict.
+
+    ``judgement`` is ``undecided`` in the shipped example, and deliberately not ``survived``: the
+    one concept that could have carried a refutation was unreadable, so the claim was never
+    actually attacked. Returning ``survived`` there would convert a gap in the knowledge base into
+    support for the claim — the inversion this role exists to prevent.
+    """
+
+    claim: str
+    judgement: str
+    #: Who refuted the claim, or ``None`` when nobody did. A refutation that names no refuter is
+    #: not a refutation (the ``verified``-without-an-actor rule, one layer up).
+    refuter: str | None
+    concepts: tuple[FalsificationConcept, ...]
+
+
+def load_falsification_example() -> FalsificationExample:
+    """Read the falsification skill's worked example. Fail-fast: a missing file raises
+    ``FileNotFoundError`` and a missing key raises ``KeyError`` (required input, contrast the
+    tolerant Step-7 verdict inbox). Resolves the path at CALL time."""
+    data = json.loads(_falsification_example_path().read_text(encoding="utf-8"))
+    return FalsificationExample(
+        claim=data["claim"],
+        judgement=data["judgement"],
+        refuter=data["refuter"],
+        concepts=tuple(
+            FalsificationConcept(
+                concept_id=c["concept_id"],
+                frontmatter_verbatim=c["frontmatter_verbatim"],
+                state=c["evidence"]["state"],
+                reason=c["evidence"]["reason"],
+                items_seen=c["evidence"]["items_seen"],
+                trust_tier=c["derived"]["trust_tier"],
+                adjudication=c["derived"]["adjudication"],
+                relied_on=c["relied_on"],
+            )
+            for c in data["concepts"]
+        ),
+    )