bug: MemWAL LSM fuzzy FTS expands max_expansions per source instead of over the merged vocabulary #9883

Problem

The MemWAL LSM FTS planner expands a fuzzy Match independently in every source (base table, SSTable generations, frozen and active memtables). Each source spends the leaf's max_expansions budget over its own vocabulary, while a single merged index spends it once over the union of every segment's vocabulary (final_query_tokens). When the cap binds, the same row can match or not depending on which tier it lives in, so results can shift as rows flush or compact.

Example — "alpha beta", fuzziness 1, AND, max_expansions = 2:

A merged index over the union returns no rows. Fuzzy OR on main has the same divergence; #9526 extends it to fuzzy AND and documents it.

A source's vocabulary is a subset of the union, so each source's own expansion is a superset of its share of the canonical set: sources only ever over-match (with LIMIT, an over-matched row can displace a canonical one from the top-k).

Proposed fix (contained in mem_wal)

1. In the LSM planner, compute each fuzzy leaf's canonical vocabulary: merge every source's own expansion per query position, then spend the budget once in position order, lexicographically. Base/SSTable expansions come from InvertedIndex::expand_fuzzy_tokens; memtable expansions from FtsMemIndex.
2. Pass the canonical groups directly to memtable arms.
3. For a base/SSTable arm whose own expansion differs from its share, rewrite the leaf into per-position single-term fuzzy Match leaves under a Boolean (must for AND, should for OR), each with max_expansions set to that position's share. The share is a lexicographic prefix of that arm's candidates, so the arm reselects exactly the canonical terms.

Open questions: score parity of the rewritten Boolean leaf versus a grouped Match, and the added planning latency (expansion becomes a barrier before dispatch; only fuzzy queries pay it).