feat: add pinned xRouteBench ingestion

Audit and pin the upstream dataset, add integrity metadata and schema
snapshots, implement value-preserving Parquet ingestion, and document the
unresolved redistribution license.

Author: MaharshPatelX

README.md

@@ -4,9 +4,9 @@
 
 Research framework for routing prompts across evolving LLM portfolios with a compact Liquid AI encoder.
 
-> **Research status:** Pre-alpha. The repository foundation is complete and the xRouteBench source audit is next. No benchmark results or trained router checkpoint are claimed yet.
+> **Research status:** Pre-alpha. The Part 2 xRouteBench audit and ingestion implementation is complete and awaiting review/merge. No benchmark results or trained router checkpoint are claimed yet.
 >
-> **Project progress:** [2 of 17 core parts complete — view the tracker](docs/PROJECT_TRACKER.md)
+> **Project progress:** [2 of 17 core parts merged; Part 2 implementation complete — view the tracker](docs/PROJECT_TRACKER.md)
 
 ## Research question
 

data/README.md

@@ -12,3 +12,19 @@ data/cache/        Regenerable local caches
 ```
 
 Do not place credentials, private prompts, or restricted data in tracked files.
+
+## xRouteBench
+
+Part 2 pins xRouteBench in `data/manifests/xroutebench.json`. Download and
+verify only its smallest sample with:
+
+```bash
+python scripts/download_xroutebench.py
+```
+
+Use `--all` only when all 47.38 MB of pinned Parquet files are needed. Both
+modes write beneath ignored `data/raw/xroutebench/<revision>/` paths.
+
+Because the pinned source declares no dataset license, do not move real rows
+into tracked fixtures. `data/fixtures/xroutebench/` contains synthetic test
+records only.

data/fixtures/xroutebench/README.md

@@ -0,0 +1,9 @@
+# Synthetic xRouteBench Fixture
+
+`synthetic_routing_sample.jsonl` is authored specifically for repository tests.
+It follows one observed xRouteBench routing schema but contains no copied
+upstream prompts, responses, model IDs, or annotations.
+
+The real xRouteBench files remain under ignored `data/raw/` paths. They must not
+be committed because the pinned upstream revision does not declare a dataset
+license or include a license file.

data/fixtures/xroutebench/synthetic_routing_sample.jsonl

@@ -0,0 +1,3 @@
+{"task_name":"synthetic_math","query":"Synthetic fixture: what is two plus two?","ground_truth":"4","metric":"exact_match","choices":null,"task_id":null,"model_name":"synthetic-model-a","response":"4","token_num":11,"input_tokens":9,"output_tokens":2,"response_time":0.01,"performance":1.0,"embedding_id":0}
+{"task_name":"synthetic_choice","query":"Synthetic fixture: select option B.","ground_truth":"B","metric":"exact_match","choices":"{\"text\":[\"alpha\",\"beta\"],\"labels\":[\"A\",\"B\"]}","task_id":"fixture-2","model_name":"synthetic-model-b","response":"B","token_num":14,"input_tokens":12,"output_tokens":2,"response_time":0.02,"performance":1.0,"embedding_id":1}
+{"task_name":"synthetic_text","query":"Synthetic fixture: return a short greeting.","ground_truth":"hello","metric":"exact_match","choices":null,"task_id":"fixture-3","model_name":"synthetic-model-a","response":"hello","token_num":13,"input_tokens":11,"output_tokens":2,"response_time":0.015,"performance":1.0,"embedding_id":2}

data/manifests/README.md

@@ -1,5 +1,11 @@
 # Dataset Manifests
 
-This directory will hold machine-readable source revisions, file metadata, schema snapshots, checksums, license notes, and split manifests.
+This directory holds machine-readable source revisions, file metadata, schema
+snapshots, checksums, license notes, and split manifests.
 
-No manifest has been created yet because xRouteBench has not completed the required source and license audit.
+- `xroutebench.json` pins the immutable source revision and all 41 files.
+- `xroutebench_schema.json` records all 17 config schemas, split row counts,
+  and non-zero null counts.
+
+The human-readable evidence and decisions are in
+[`reports/xroutebench_audit.md`](../../reports/xroutebench_audit.md).