Declare causal retrieval attention contract

README.md

@@ -170,8 +170,9 @@ pip install causal-conv1d flash-linear-attention
 ```
 
 - `causal-conv1d` + `flash-linear-attention` are required (the hybrid layers import them at runtime).
-- Attention must be **SDPA**. Retrieval runs bidirectional attention on the full-attention layers;
-  `flash_attention_2` silently ignores the 2-D mask and scores as if causal.
+- This merged checkpoint was selected and evaluated with **causal** full-attention layers. Load its
+  `retrieval_attention_contract="causal"` metadata without overriding it.
+- Use **SDPA**, matching the evaluation runtime of record.
 
 ## Usage
 
@@ -206,20 +207,20 @@ print(scores.shape)  # torch.Size([2, 2])
 ```
 
 `config.json` carries `dim=320`, so `custom_text_proj` is sized correctly at load, with no manual
-config edits needed.
+config edits needed. It also carries `retrieval_attention_contract="causal"`; a conflicting runtime
+override is an error.
 
 ## Serving with vLLM
 
 vLLM serves this model natively through its pooling runner (the `ColQwen3_5` architecture), returning
 the per-token multi-vectors for late-interaction scoring. It requires a vLLM build that includes the
-ColQwen3.5 retrieval-correctness fix ([vllm-project/vllm#46108](https://github.com/vllm-project/vllm/pull/46108),
-merged 2026-06-22): build from `main`, or use a release tagged after that date. The fix runs the
-backbone bidirectionally and restores the projection bias, so vLLM reproduces the `transformers`
-reference within run-to-run noise. The server uses the stock chat/image processor, so the ColQwen3.5
+projection-bias fix and explicit retrieval-attention metadata support. The handler reads
+`retrieval_attention_contract="causal"` from this checkpoint and constructs all eight full-attention
+layers as `AttentionType.DECODER`; it must reject missing, conflicting, or unsupported contract
+metadata. The server uses the stock chat/image processor, so the ColQwen3.5
 prompt contract is applied **client-side**: wrap each page image in the instruction template, append
 the query-augmentation tokens to each query, and set the visual-token budget through
-`mm-processor-kwargs`. Prefix caching and chunked prefill must be **off** (bidirectional attention and
-the GatedDeltaNet hybrid both break the causal-prefix invariant).
+`mm-processor-kwargs`. Prefix caching and chunked prefill remain off for the GatedDeltaNet hybrid.
 
 ```python
 import torch

config.json

@@ -4,6 +4,7 @@
   ],
   "image_token_id": 248056,
   "model_type": "qwen3_5",
+  "retrieval_attention_contract": "causal",
   "text_config": {
     "attention_bias": false,
     "attention_dropout": 0.0,