refactor: make corpus operations functional [AAI-766]

[ypicard]

Superseded by platform PR #5898, which keeps input preservation in e14c and uses released PaperQA. The platform change requires no PaperQA or server API migration.

Avoid whole-corpus copies by returning updated documents without changing inputs.

Breaking API change: add, retrieve, evidence, and query operations return (docs, result); clear, delete, and merge_docs return docs. Callers retain the returned documents to reuse embeddings and indexes. Query operations also leave input sessions unchanged. Use the standalone merge_docs(*corpora) function to combine independent results and retain the first corpus’s retrieval state.

Qdrant corpus retrieval raises NotImplementedError; use NumPy. Direct Qdrant operations remain available. Custom vector stores must implement fork().

Corpus benchmark: 2,000 chunks, 256-dimensional embeddings, median of 5 runs. Baseline includes the old deep copy. Memory means peak traced allocations, not RSS; excludes network and LLM calls.

Operation	Before (ms)	After (ms)	Before (MiB)	After (MiB)
Insert	143.100	0.034	10.074	0.023
Cold retrieval	153.023	13.032	13.403	8.090
Warm retrieval	152.004	3.378	13.363	4.053

[ypicard]

👀