test(core): add reranker real-model smoke and quality harness

[phernandez]

Phases 2-3 of #1231 — the remaining automated items. Test-only: six files under
test-int/semantic/, zero production changes.

What's added

• Real-model smoke (test_real_fastembed_reranker.py, semantic-marked): loads the
  actual jinaai/jina-reranker-v1-tiny-en cross-encoder through the production
  provider/factory path (no stubs), asserts the relevant doc wins with calibrated [0,1]
  scores and deterministic ordering, plus one end-to-end hybrid search through an indexed
  repository with reranking enabled.

• Quality regression harness (test_reranker_quality.py + ranking corpus/metrics
  extensions): golden ranking-sensitive queries measured reranker-off vs reranker-on with
  real embeddings + real cross-encoder. Asserts on >= off for hit@1/MRR plus an absolute
  hit@5 floor so a silent rerank no-op cannot pass. Measured on the curated set:

  Mode	hit@1	hit@5	MRR
  Reranking off	0.833	1.000	0.917
  Reranking on	1.000	1.000	1.000

  No queries needed exclusion; the real tiny model degraded nothing.

• Latency benchmark (test_reranker_latency.py, semantic+benchmark-marked, report-
  focused with a generous catastrophic-regression ceiling only). Warmed local numbers,
  20 iterations:

  Search	P50	P95
  Vector off	9.5 ms	10.1 ms
  Vector on	102.5 ms	238.4 ms
  Hybrid off	13.2 ms	17.6 ms
  Hybrid on	98.2 ms	463.8 ms

  ~90 ms P50 overhead for the local tiny cross-encoder — comfortably inside the latency
  headroom that motivated Add a rerank stage to search: ~half of LoCoMo benchmark misses are ranking failures, and there is ~20x latency headroom #950.

Verification

• The three new test files: pass, run twice for determinism (16s total once models are
  cached).
• just typecheck, just lint: clean.
• Note: test_semantic_quality[asyncio-postgres-fastembed] can exceed the 120s per-test
  timeout on a loaded local machine — reproduced identically on main, pre-existing and
  unrelated to this PR.

Ticks the Phase 2 fastembed-smoke and both Phase 3 boxes on #1231. Remaining there: the
deferred live LiteLLM run and the optional live smoke. These numbers also set the baseline
for revisiting #951 (entity boosting) and closing #618.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CPdSXDbYyhyZ1TwgFnpEv8