test(core): close reranker phase-1 coverage gaps

[phernandez]

Phase 1 + the small-fixes section of #1231 (reranker test plan, post-#1143 audit).

Commits

• fix(core): align reranker defaults and timeout config — the FastEmbed constructor's
  stale Xenova/ms-marco-MiniLM-L-6-v2 default now references the shared
  DEFAULT_FASTEMBED_RERANK_MODEL constant; new reranker_timeout config field
  (default 30.0, gt=0) threads through the factory to the LiteLLM provider and joins the
  singleton cache key (litellm only, so fastembed keys are unaffected). Documented in the
  semantic-search guide; FastEmbed ignores it by design.
• fix(mcp): preserve retryable search outages — the ChatGPT-compatible search adapter
  no longer collapses rerank 503s into "Internal search error": it detects search_notes'
  service-unavailable response and returns a distinguishable "Search temporarily
  unavailable, retry shortly" payload so OpenAI clients can retry.
• test(core): close reranker phase-1 coverage gaps —
  • The 9 repository-integration rerank tests no longer skip on Postgres: a backend-aware
    rerank_search_repository fixture runs the identical scenarios against
    SQLiteSearchRepository and PostgresSearchRepository (37 pass on each backend, zero
    skips), so hybrid/vector rerank ordering is guarded in the main unit shards.
  • New FTS negative guard: with a provider configured (exploding fake), text/title/
    permalink searches never invoke it and match baseline ranking.
  • Wiring-seam tests: create_search_repository injects the provider into both backends;
    repository __init__ self-resolution only fires when semantic_search_enabled.
  • reranker_max_document_chars verified end-to-end from config to the document text the
    provider receives.

Verification

• Focused suites (pipeline, factory, providers, ChatGPT adapter, config CLI): 148 passed.
• BASIC_MEMORY_TEST_POSTGRES=1 pipeline run: 37 passed, no skips (testcontainers).
• just typecheck, just lint: clean.

Ticks the Phase 1 and small-fixes checkboxes on #1231; Phases 2-4 (real-model smoke, live
provider harness, quality regression, operational drills) remain tracked there.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CPdSXDbYyhyZ1TwgFnpEv8