docs: implementation plan for backend breadth and cross-family RRF blending (#52)

[TKaltofen]

Summary

Planning document only, no code: docs/plan-issue-52-backend-breadth.md lays out the implementation plan for issue #52 (backend breadth for the existing six connector families plus cross-family RRF blending). Refs #52.

The plan covers:

• Per-family backends with class names, selector values, extras, and CI treatment: sentence-transformers CrossEncoder (rerank), LlamaIndex keyword-table retriever path and txtai keyword mode (orchestrator), llama.cpp via Llama.from_pretrained plus an optional claude_cli adapter (generate), and an explicit optional decision point for retrieve late interaction.
• Cross-family RRF blending on doc_id via the existing rrf_fuse in connectors/fusion.py, with a demo hook in cli/swap_demo.py and a test list, per the deferred design decided in chore(deps): bump the all-dependencies group with 4 updates #46.
• Honest-surface narrowing rows, extras table, PR slicing and order, definition-of-done mapping, and open questions.
• Ordering rule: implementation waits for chore(deps): bump urllib3 from 2.6.3 to 2.7.0 #51 (eval harness) so each backend lands with a metrics comparison row.

Verification

• tox fully green (692 passed, 7 skipped; ruff format, ruff check, mypy --strict, bandit).
• Dual independent review of the doc: a deep Claude review (5 findings, all fixed in the second commit: conftest marker reuse, llama-cpp-python acquisition path, claude_cli adapter framing, survey-row wording, LlamaIndex wrapped-surface wording) and a codex review (no findings).

[TKaltofen]

Opened against the wrong base repo; this planning PR belongs on the TomKaltofen fork where issue #52 is tracked. Recreating there.