feat(phase-61): admin backfill-index (61-01)

[RichardHightower]

Closes the pre-v3.1 empty text_preview gap (#41 / OPS-01). Stacked on #49 (needs the sequence-0 checkpoint fix).

What

memory-daemon stop
memory-daemon admin backfill-index --index bm25 --from-sequence 0
memory-daemon start --foreground

• Default start resumes from the index checkpoint (second run = 0 new).
• --from-sequence 0 replays the event log so stores whose outbox is already drained still get text into Tantivy (idempotent upsert by event_id).
• --dry-run prints counts, writes nothing (checkpoint unchanged).
• Refuses if the daemon PID is live; the error names memory-daemon stop.
• Progress n/total on stderr every batch. max_batches=1 in tests is the SIGINT-after-batch-1 equivalent.

This is not admin rebuild-indexes (TOC/grips) and not admin rebuild-bm25 (a prune). Vector backfill of raw events is still a no-op without grips; use --index bm25 for this gap. Import.rs was not cherry-picked — it is not on this line.

A v3.0 RocksDB dump at 68ab122 is not in-tree (impractical). The e2e seeds 50 events with real OutboxEntry JSON, never runs the indexer, then backfills.

Verify (targeted; same bar as #49 / #50)

cargo fmt --check
cargo clippy -p memory-storage -p memory-indexing -p memory-daemon --all-targets -- -D warnings
cargo test -p memory-storage --lib -- test_get_events_page test_count_outbox_from
cargo test -p memory-indexing --lib -- backfill
cargo test -p memory-daemon --lib -- backfill refuse
cargo test -p e2e-tests --test backfill_test

All of the above are green locally.

Honest caveats

• --index all / --index vector loads Candle. The OPS-01 e2e uses --index bm25.
• Event-log fallback assigns synthetic sequences 0..N-1 and pins the checkpoint to max(last, outbox_head-1) so a later live drain does not re-process.