fix(undo): capture version_id inline to avoid cascade-artifact pollution

[mfreed]

ExecuteUndoTransaction restores a set of affected files in three phases: Step 1 DELETEs rows that were created after the undo target point; Step 2 UPSERTs rows back to a history-snapshot state for edits/renames/ deletes; Step 3 inserts a new type='undo' log entry per affected file, each carrying a version_id pointer at the history row written by Step 1/Step 2's mutation. That pointer is what a future undo-of-undo dispatches on -- it must resolve to the row written by the actual mutation (tombstone for a recreate, rename/edit/delete OLD-state row for the others), not to any other row in this file_id's history.

Step 3 captured that version_id by querying the history table for each file's "newest row" -- SELECT version_id FROM WHERE file_id= $1 ORDER BY version_id DESC LIMIT 1 -- AFTER all Step 1/2 mutations were done. That post-hoc lookup is wrong when Step 2 touches a parent dir: the bump_parent_mtime AFTER trigger cascades on every child restore, firing the archive trigger again on the parent and writing a no-semantic-content edit row. That cascade row becomes "newest" for the parent's file_id, so Step 3 records IT as the parent's undo-log version_id instead of the parent's own restore snapshot. A subsequent undo-of-undo restores from a snapshot whose filename matches current state -- a no-op. The user-visible symptom: roll forward after an undo-to-savepoint silently drops a directory rename while correctly re-applying child renames in the same batch.

Move the version_id capture inside the Step 1 (DELETE) and Step 2 (UPSERT) loops, immediately after each file's own mutation, before control returns to ExecuteUndoTransaction. At that instant the file's own snapshot is guaranteed-newest in its history (row-level BEFORE trigger, one row per UPSERT/DELETE, no other SQL between mutation and capture, PG18 per-session monotonic uuidv7). Stash the captured version_id in a per-file map; Step 3 reads from the map instead of re-querying.

Correct regardless of iteration order, cascade direction, or future triggers added on the source table -- the invariant is local to each iteration rather than depending on the cumulative state of history at end-of-transaction.

Adds TestSynth_UndoOfUndo_DirRenameWithChild (integration, reproduces the user's reported bug end-to-end) and TestExecuteUndoTransaction_CapturesVersionIDInline_NotPostHoc (unit, exercises the cascade with seeded triggers). Both verified to fail pre-fix and pass post-fix.