Recover misplaced joint blocks and preserve unchanged entity citations

docs/joint-case-planning-v2.md

@@ -122,3 +122,33 @@ verification is not silently filled: the scoped statement correction must provid
 The original 2026-10-20 replies and a bounded validation context are preserved in the
 curated rejection corpus. Offline regression checks hashes, reference union, sibling
 preservation and rejection of invented references. No paid API request was needed.
+
+
+## Structural recovery revision 2
+
+New contexts freeze `recovery_version = 2`; revision 1 and unmarked historical
+contexts keep their original request/projection behavior. This revision includes
+revision 1 dependency completion and sibling preservation.
+
+One exact misplaced wrapper is recoverable: a sole top-level `hypotheses` object
+containing exactly an `interpretation` object and a `hypotheses` array. These are
+copied to the two intended blocks, without changing or dropping any content.
+Competing top-level interpretations (even null), extra sibling keys and malformed
+wrappers are never guessed away. Both recovered blocks still undergo all existing
+schema, source, dependency and semantic contract checks. The original call stays
+unchanged and remains the usage/provenance source during replay and import.
+
+During a scoped entity correction, existing eligible identity/mention citations
+are restored only when BOTH entity_id and reason are unchanged. Eligible references
+come from the original statement and the frozen entity's matching capture/mention
+provenance. New evidence is never invented. Changed claims receive no restoration;
+unknown replacement sources remain present and invalid. Requests explicitly explain
+which existing citations to preserve. When no reusable hypotheses exist, the request
+explicitly requires an empty reviews array instead of asking to reconstruct them.
+
+The curated `joint-nested-envelope` corpus preserves the actual rejected replies,
+their hashes and bounded context. Offline tests confirm that the initial reply passes
+full interpretation and hypothesis validation after lossless wrapper recovery and
+existing dependency normalization. A separate regression confirms the later scoped
+patch passes interpretation validation after preserving unchanged identity evidence.
+No paid API request was executed to establish these results.