Merge DGR-025 (certified-artifact-byte recipe identity) from ralph-fable-loop lane

scripts/gen_recipe_fingerprint_vectors.py

@@ -34,11 +34,25 @@ from meshnet_node.runtime_recipe import (  # noqa: E402
     DerivativeBinding,
     RuntimeRecipe,
     ShardIdentity,
+    tokenizer_identity,
+)
+from meshnet_tracker.recipe import (  # noqa: E402
+    parse_identity,
+    tokenizer_identity as tracker_tokenizer_identity,
 )
-from meshnet_tracker.recipe import parse_identity  # noqa: E402
 
 VECTORS = _ROOT / "tests" / "data" / "recipe_fingerprint_vectors.json"
-SCHEMA_VERSION = 1
+SCHEMA_VERSION = 2
+
+# The tokenizer axis is content-addressed: a digest over the tokenizer's actual
+# bytes, never a repository label. These example bytes are part of the wire
+# contract exactly like the digests derived from them.
+_TOKENIZER_FILES = {
+    "tokenizer.json": b'{"version":"example","vocab":{"a":0,"b":1}}\n',
+    "tokenizer_config.json": b'{"model_max_length":8}\n',
+}
+_TOKENIZER_IDENTITY = tokenizer_identity(_TOKENIZER_FILES)
+assert _TOKENIZER_IDENTITY == tracker_tokenizer_identity(_TOKENIZER_FILES)
 
 _RECIPE = RuntimeRecipe(
     weight_quantization="Q4_K_M",
@@ -46,10 +60,13 @@ _RECIPE = RuntimeRecipe(
     compute_dtype="float32",
     kv_dtype="q8_0",
     kv_layout="paged-v1",
-    tokenizer_revision="0123456789abcdef",
+    tokenizer_revision=_TOKENIZER_IDENTITY,
     architecture_adapter="llama/range-v1",
     backend_id="llama.cpp",
-    runtime_version="llama.cpp@" + "d" * 40 + "+patchstack." + "e" * 64,
+    runtime_version=(
+        "llama.cpp@" + "d" * 40 + "+patchstack." + "e" * 64
+        + "+build." + "f" * 64 + "+artifact." + "a" * 64
+    ),
     recipe_id="example-gguf",
     recipe_version="1",
     catalogue_version="2026.07.1",