Replaces the old no-op bulkdelete (which discarded callback results)
with a correctness fix that:
1. Spills memtable to segments
2. Identifies segments containing dead CTIDs (O(segments) memory)
3. Rebuilds affected segments from heap via TpBuildContext
4. Updates metapage statistics

This prevents false positives when Postgres reuses CTIDs after VACUUM
frees heap line pointers.

All four test scenarios pass:
- CTID reuse: alpha_matches=0 (fix confirmed)
- All-deleted segment removal works
- Multi-level vacuum preserves remaining docs
- No-op vacuum path works

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The original test used WHERE <@> < 0 which triggers a seqscan that
re-evaluates the operator on the heap tuple, masking the bug. The
fix uses ORDER BY <@> LIMIT which triggers an index scan where
xs_recheck=false causes Postgres to trust the stale index entries.

Verified: test produces alpha_false_positives=1 without the fix,
and alpha_false_positives=0 with the fix.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

DatumGetTextPP may detoast, allocating memory. By calling it inside
per_doc_ctx, the detoasted datum is freed on each MemoryContextReset
instead of accumulating in the outer context.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Replace the complex alpha_false_positives subquery pattern with
a direct SELECT that expects 0 rows. The index scan returns nothing
for deleted terms when the fix is working correctly.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>