feat(verdicts): a cross-base candidate collision is reported, never dropped in silence

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

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

src/portfolio_optimiser/verdicts.py

@@ -110,6 +110,27 @@ def verdict_key(features: ProposalFeatures) -> str:
     return _mint_id(features)
 
 
+@dataclass(frozen=True)
+class VerdictCollision:
+    """Two knowledge bases described ONE candidate, and the second verdict was dropped.
+
+    ``VerdictStore.add`` is first-write-wins per id and ``_mint_id`` excludes the corpus BY
+    CONSTRUCTION, so this drop is CORRECT — one candidate, one learning key — but it used to be
+    silent, and a base whose finding never entered the store looked exactly like a base that found
+    nothing (operator decision D2).
+
+    A DIAGNOSTIC, never an aggregate over bases: it reports that two bases described one candidate
+    and never merges, ranks or sums them, which is why it needs no combination rule. ``add`` itself
+    is unchanged — it is called from inside ``run_project``/``run_portfolio``, where a drop can
+    equally be against a Step-7 inbox verdict or a bundle seed, so only the dispatcher holds the
+    id -> base map that makes "a SECOND base" a statement worth making.
+    """
+
+    verdict_id: str
+    first_bundle_id: str
+    second_bundle_id: str
+
+
 def capture_verdict(features: ProposalFeatures, decision: str, rationale: str) -> Verdict:
     """Layer-2 out-of-band verdict constructor: mint a stable content-hash id (the
     learning-loop key) and build the ``Verdict`` to persist in the store."""