feat(rerank): run /v1/rerank on mlx-swift-lm's MLXRerankers

[magicnight]

Replaces the hand-written BERT cross-encoder behind /v1/rerank with mlx-swift-lm 3.32.3's MLXRerankers, and verifies the endpoint against real checkpoints for the first time.

What changes

• RerankEngine holds an architecture-neutral RerankerContainer from RerankerModelFactory; CrossEncoderModel and the pair-assembly code are deleted. The factory reads config.json and picks encoder cross-encoders (BERT / RoBERTa / XLM-RoBERTa with a sequence-classification head), Qwen3 causal rerankers (yes/no logit margin) or Jina reranker v3 (listwise).
• load passes allowUnverifiedModel: true on purpose: upstream only accepts identifiers containing "rerank", which would refuse cross-encoder/ms-marco-MiniLM-L-6-v2. The format decision is ours, made once in ModelLibraryManager.upgradeFormat on the same evidence plus the single-logit head check.
• Detection learns two shapes: JinaForRanking is a reranker outright; a qwen3 + Qwen3ForCausalLM config (byte-identical to a chat model's) is one when the directory name or Hub repo id contains "rerank", upstream's own rule. The cache scan now hands the repo id to upgradeFormat.
• /v1/rerank exposes the model's own score: unchanged for BERT rerankers (sigmoid of the single logit), normalized 0...1 for Qwen3, cosine similarity for Jina v3. rankAndTruncate / rerankResults are generic over the score type; rerankSigmoid is gone. Request and response shapes are untouched.
• Pairs are micro-batched by upstream (16 pairs or 8,192 token slots per forward pass) and truncated to the model context.

Verification

• Pure suites: 961 tests / 101 suites green under swift test; the detection suite gained five tests (name rule, case-insensitivity, scope, Jina, HF-cache repo id). Four mutations of the new rules each turn exactly one test red.
• Gated real-checkpoint smokes (MACMLX_RUN_RERANK_SMOKE=1, xcodebuild):
  • cross-encoder/ms-marco-MiniLM-L-6-v2 vs a PyTorch fp32 reference (fixture + docs/reference/capture_ms_marco_reranker.py): logits 8.820 / −11.242 / 9.565 / −11.075 against 8.846 / −11.246 / 9.585 / −11.074, worst |Δ| 0.025 (0.3 %), identical ordering. Tolerance in the test is 0.05, twice the measured worst.
  • mlx-community/Qwen3-Reranker-0.6B-4bit: scores 0.988 / 5e-5 / 0.990 / 9e-6, the two answering documents first.
  • End to end through POST /v1/rerank with the encoder: 200, top_n honored, return_documents echoes by original index, descending order.
• Not run: Jina reranker v3 (no checkpoint pulled); multi-label encoder heads still route to the embedder path, as before.

Review round (second commit)

• Detection narrowed to what MLXRerankers can actually score: a qwen3 + Qwen3ForCausalLM checkpoint is a reranker only when its repo name reads Qwen3-Reranker and, if it ships 1_LogitScore/config.json, that file declares both the true and the false token. zerank-2-reranker and ctxl-rerank-v2 (same config, single-token protocol) stay chat-typed and keep answering 400. JinaForRanking is skipped when the config declares sliding-window layers (Jina v3.5). Six more detection tests; six mutations each turn exactly one test red.
• A blank query or document is a 400 before the model is resolved (v0.9.0 scored an empty document; the new engine would have thrown a 500 after a cold load). Request-shaped upstream errors map to 400 through RerankRequestError. Two route tests prove the 400 precedes the load.
• Ties rank by ascending index, as upstream does.
• The 0.025 logit drift was TF32, not the model: with MLX_ENABLE_TF32=0 ms-marco matches PyTorch to 1.9e-6. The parity test gates at 1e-4 with TF32 off and 0.05 under the M5 default, and prints which mode ran; both modes pass on this machine.
• CHANGELOG entry merged into the existing Changed section; BGE and Jina stated as not run; capture script pins the checkpoint revision and records library versions; stale comments about the deleted head rewritten.