Introduce kind: text requests for BM25 over a text index

[agourlay]

Adds BM25 over a payload field's text index to the config path, the Query.text variant Qdrant dev added for 1.19.3 (qdrant/qdrant#10786 and the series before it).

Builds against qdrant/rust-client's dev branch, which carries the new API since qdrant/rust-client#299 merged (71143d8).

Search config: kind: text

requests:
  - kind: text
    using: body                    # payload field with a scoring text index
    k: 1.2                         # optional, BM25 term-frequency saturation
    b: 0.75                        # optional, BM25 length normalization, 0..1
    filters: [...]                 # optional, as for other kinds
    source:                        # random words, or a file of queries
      type: random                 # random | file
      vocab_size: 20000
      min_length: 2
      max_length: 5
      distribution: zipf

• random draws word_<id> words, the same vocabulary text payloads are generated from, so the queries hit the corpus.
• file reads one query per line (path, strategy: from-start | random-sample), loaded once at startup. Blank lines are skipped, and an empty file fails at startup.
• using names the payload field and is required. k must be >= 0 and b within [0, 1], the same bounds the server validates.
• --search-quality is refused with a text request, because BM25 over a text index is exact and there's no exact reference to compare against. --prefetch is refused too, since there's no vector to rescore. --acorn works on the request's filters.

Upload config: text fields

fields:
  - name: body
    type: text
    tokenizer: word
    scoring: bm25                  # TextIndexParams.scoring; implies phrase matching
    lowercase: true                # optional
    stopwords: english             # optional, stopword language
    stemmer: english               # optional, Snowball stemmer language

scoring, lowercase, stopwords and stemmer apply only to type: text fields with index: true, and are refused elsewhere so they can't be silently dropped.

Other changes

• create_request_builder takes the query as a single MainQuery (a vector or a text query) instead of a vector plus a text argument, which keeps it under clippy's argument limit.
• New examples upload-bm25-text and search-bm25-text, plus a README section.
• The schema reference example gains a text field carrying the new options.

Verification

• cargo test --release: 207 passed, 7 of them new (config parsing and validation, random and file query sources, request building, flag guards, text index options).
• cargo clippy --release --all-targets -- -D warnings: clean. cargo fmt --check: clean.
• End to end against Qdrant dev (850859ec9): upload, random-source and file-source searches all pass, with the server reporting the BM25-scoring index. See the comment below.
• Re-run after moving to the client's dev branch: 207 tests, clippy and fmt pass. 71143d8's tree is identical to the d960517 the change was built against.

🤖 Generated with Claude Code

[agourlay]

End-to-end check against Qdrant dev at 850859ec9 (1.19.3-dev, --profile perf), whose proto is identical to rust-client#299's. Run on a laptop for correctness, not as a benchmark.

• bfb upload --example upload-bm25-text -n 20000: 20,000 points uploaded, collection green. The server reports body as {"type": "text", "tokenizer": "word", "phrase_matching": true, "scoring": {"type": "bm25"}}, with phrase matching turned on by scoring as documented.
• bfb search --example search-bm25-text -n 5000 --search-limit 10 -p 4: no errors, median 688 q/s, server p99 40 ms.
• kind: text from a type: file source (strategy: from-start, k: 2.0, b: 0.0): no errors, median 1,244 q/s.
• A direct REST {"query": {"text": "word_1 word_42"}, "using": "body"} on the same collection returns ranked points with BM25 sums (2.677, 2.638, 2.576 for the top three).
• No ERROR lines in the server log.

The Client version 1.16.1-dev is not compatible with server version 1.19.3-dev warning comes from the client crate's version string on the dev line; the run is unaffected.

🤖 Generated with Claude Code