feat(phase-60): live-backend isolation for LOCOMO (60-01)

[RichardHightower]

Summary

Phase 60-01 of v3.2 Prove It (#46). --backend cli no longer bleeds events across LOCOMO conversations.

• Spawn one daemon per conversation (tempdir, free port, --pid-file)
• Drain waits poll GetIndexCheckpoints (no thread::sleep stand-in)
• Result JSON records isolation: per-conversation daemon and per-conversation drain_wait_ms
• --limit-questions ready for the 60-02 dry-run
• CI job bench-cli-smoke runs the 1-conversation fixture against a spawned daemon

Two bugs that would have made the live number 0.0:

1. Query-path Tantivy reader never reloaded after the indexer committed
2. Outbox sequence 0 did not advance the BM25 checkpoint, so a one-event drain hung

Closes BENCH-10 / BENCH-11. Unblocks #39 (60-02, maintainer-owned).

Acceptance

• memory-bench locomo --backend cli --scorer mock --dataset benchmarks/fixtures/locomo-smoke.json → isolation: per-conversation daemon + drain_wait_ms
• cli_isolated_daemons_do_not_bleed (under MEMORY_BENCH_LIVE=1)
• CI job bench-cli-smoke added
• No std::thread::sleep in the cli-backend path (poll_pause uses recv_timeout)

Verified locally: locomo smoke drain_wait_ms=50390, bleed test 65s, two daemons, no cross-talk.

Test plan

• cargo fmt --all -- --check
• cargo clippy -p memory-bench -p memory-indexing --all-targets -- -D warnings
• CI bench-cli-smoke + workspace test/clippy/doc