test: cover search no-Lance error path + multi-list RRF (#358)

[HumanBean17]

What

Closes the two highest-value coverage gaps from #358. (No production code changed — tests only.)

search no-Lance error path (the #1 gap)

Every search test monkeypatched run_search, so search_v2's genuine except Exception envelope (mcp_v2.py) was never exercised in CI — a regression turning the graceful failure into a traceback would ship green. New test_search_no_lance_index_returns_failure_envelope runs search against an index dir with no Lance tables and no run_search monkeypatch, and asserts:

• success is False with a non-empty message and no traceback, and
• find still succeeds against the same graph (the failure is vector-specific, not a crash).

multi-list RRF

_rrf_merge was covered only for the weighted two-list case. New test_rrf_merge_reinforced_row_across_lists_outranks_singleton asserts a row reinforced across two ranked lists accumulates score and outranks a singleton, dedups by (filename, range_start, range_end), and orders by summed score.

Already covered / deferred

• resolve no-match is already covered on master (test_resolve_status_none_returns_nonempty_message) — the issue was stale on this point.
• Installer config-schema migration: deferred — handle_rerun has no explicit config-version marker to migrate on, so a test would be speculative/overfit. Worth revisiting if/when a config_version key is introduced.
• Promoting the vectors flow / concurrent-optimize race guard into CI: deferred — that's a model-downloading CI policy decision (the tests are JAVA_CODEBASE_RAG_RUN_HEAVY-gated), better made by the maintainer than bundled here.

.venv/bin/python -m pytest tests/test_mcp_v2.py tests/test_search_lancedb.py -q   # 112 passed, 1 skipped

Addresses the top-2 highest-value coverage gaps from #358 (search no-Lance error path; multi-list RRF). resolve no-match was already covered on master; installer config-schema migration and CI vectors-flow promotion remain open (see Already covered / deferred above), so this PR intentionally does not auto-close #358.

🤖 Generated with Claude Code

[HumanBean17]

Addressed 2 review findings:

1. E303 — the two new tests were each followed by 3 blank lines (max 2). Collapsed to 2. (Latent: the repo ruff config doesn't enable E3.)
2. Closes #358 auto-close — the body would auto-close the issue on merge while 2 acceptance criteria are deferred. Replaced with non-closing wording ("Addresses the top-2 … intentionally does not auto-close test coverage & CI gaps: search no-Lance error path, resolve no-match, installer migration, ranking, heavy-gating #358").