[AIGNLAI]

Summary

• Add VL-RouterBench through the existing BaseDatasetLoader and router/evaluation contracts, with frozen metadata for 14 source datasets and 17 candidate VLMs, benchmark/task/model configs, README news and preparation documentation.
• Read named upstream Steps 1/2 JSONL/Parquet/CSV exports rather than positional matrices or pickle indexes. Align queries and candidates by identity, preserve exported split assignments, reject incomplete/invalid records by default, and record source hashes and evaluation protocol sidecars.
• Resolve TSV row-ordinal image references and reference chains, preserve original encoded image bytes, validate local paths and explicitly handle multiple images. Provide an opt-in archive downloader with a pinned SHA256, safe extraction and overwrite guards; the loader does not execute upstream preparation code or paid model calls.
• Compute explicitly labeled reference costs from frozen prices and released token statistics. Record estimated-output flags; these costs are not certified provider bills. Guard against cross-split identical text inputs, with an explicit recorded drop-query policy when needed.
• Add RouteFM-Qwen joint-embedding support via a local Qwen3-VL-Embedding-8B encoder or an explicitly configured compatible service. Retain default BGE text behavior and enforce frozen, normalized 4096-D Qwen features with no silent text-only fallback.
• Add a separately labeled ProfileRouter multimodal extension using matched CLIP text/image encoders and normalized weighted fusion in one 512-D space; retain the original text path and existing graph/cost/scoring machinery.
• Add optional top-level JSON CLI overrides, example configs, encoder/API contract tests and expanded all-method smoke coverage. Correct the CPU-fallback test expectation and include the local encoder interface dependency in CI.

Validation

• Local python -m unittest discover -s tests -q: 106 tests passed.
• The offline regression includes 176 method/benchmark/modality subcases: 30 text methods on each of four new benchmark fixtures, plus 28 compatible image methods on each MMRBenchV2/VLRouterBench fixture.
• Compileall, every JSON config and relative documentation links checked; git diff --check passed.
• Real bounded VL-RouterBench records: 30/30 text (36 train / 32 test queries) and 28/28 image (32 / 32), using RealWorldQA and the two Janus-Pro candidates. Training/context construction, prediction, allocation, metrics and JSON writing execute normally. These batch smoke runs use input-boundary surrogate features, not all production semantic encoders. RouteFM loads released BGE/Qwen router weights as appropriate; tiny Transformer fixtures and the NIRT auxiliary-annotation stub are explicitly recorded.
• Separate real pretrained CLIP ViT-B/16 ProfileRouter smoke: 4 train / 4 test RealWorldQA queries, two candidates; train/predict/evaluate passed without encoder substitution. Changing the image while holding the question fixed changed the query vector (L2 delta approximately 0.7044).
• Loopback HTTP contract tests cover actual image-byte transport, indexed response reordering, normalization and invalid-input rejection without calling a paid service.

Boundaries / not claimed

• Upstream preparation is required; the raw archive is not a tabular load_dataset source. Real records were prepared with the reviewed upstream exporter revision; regenerated split files are not independently certified original-paper partitions.
• The real-data checks above cover one source dataset and two candidates, not all 14 datasets/17 models or paper-quality reproduction.
• Genuine Qwen3-VL-Embedding-8B end-to-end routing has not been validated locally: the encoder weights/service are unavailable. The released Qwen router was smoke-tested with explicitly labeled text/image-dependent surrogate vectors. The production local/API encoder paths must be checked with the actual encoder resource before claiming that validation.
• ProfileRouter's CLIP mode is an ORBIT multimodal extension, not the original paper's multimodal reproduction.
• ModelSAT and RouteLLM_BERT remain text-only in this adapter/test policy. On VL exports without released response strings, EA-RAM's optional response-based ex-post evaluator is not fitted; responses are never fabricated.
• This PR does not certify all 30 methods across all 9 ORBIT benchmarks under default production configurations. Full cross-validation of the five legacy benchmarks remains separate work.