refactor: dedupe LLM resilience/ingestion retries, type RAG source boundary

[n0nuser]

Summary

• Kill 4 type: ignore casts on the RAG→API source boundary via SourceRef(**s) unpacking
• Collapse ResilientProvider's 4 duplicated breaker/fallback blocks into _call_with_breaker/_stream_with_breaker
• Collapse ingestion's two-pass retry loop into a shared _ingest_batch helper
• Type IngestionService.bm25_index as Bm25Index | None instead of Any | None
• Add AgentApiError so /agent/query follows the same error-mapping convention as every other router
• 404 DELETE /collections/{name} on unknown name instead of silently no-oping
• Cap concurrent ingest job submission (max_pending_ingest_jobs, default 10), reject past-cap submissions with 429

No behavior changes except the two explicitly-scoped additions (404 on unknown collection delete, 429 on job-submission cap) — everything else is pure refactor, verified against the existing test suite.

Findings came from a /thermo-nuclear-code-quality-review + python-anti-patterns + python-code-style + api-design-principles audit of the whole localrag/ package, then implemented in 3 phases by separate agents on disjoint files.

Test plan

• ruff check localrag/ tests/ — all checks passed
• mypy --ignore-missing-imports --no-strict-optional on all touched files — no issues
• pytest tests/ --ignore=tests/integration — 142 passed
• Reviewer: confirm 429/404 behavior changes are acceptable (not pure refactor)

https://claude.ai/code/session_01NFJtv7kyDmx4W8u76WW9EC