diff --git a/Cargo.lock b/Cargo.lock
index 91821e35..7df7802d 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -220,17 +220,6 @@ dependencies = [
  "syn",
 ]
 
-[[package]]
-name = "bindgen_cuda"
-version = "0.1.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "282be55fb326843bb67cccceeeaf21c961ef303f60018f9a2ab69494dad8eaf9"
-dependencies = [
- "glob",
- "num_cpus",
- "rayon",
-]
-
 [[package]]
 name = "bit-set"
 version = "0.5.3"
@@ -310,67 +299,6 @@ version = "1.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
 
-[[package]]
-name = "candle-core"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c15b675b80d994b2eadb20a4bbe434eabeb454eac3ee5e2b4cf6f147ee9be091"
-dependencies = [
- "byteorder",
- "candle-kernels",
- "candle-ug",
- "cudarc 0.19.4",
- "float8 0.6.1",
- "gemm 0.19.0",
- "half",
- "libm",
- "memmap2",
- "num-traits",
- "num_cpus",
- "rand 0.9.2",
- "rand_distr 0.5.1",
- "rayon",
- "safetensors 0.7.0",
- "thiserror 2.0.18",
- "yoke 0.8.2",
- "zip",
-]
-
-[[package]]
-name = "candle-kernels"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b8455f84bd810047c7c41216683c1020c915a9f8a740b3b0eabdd4fb2fbaa660"
-dependencies = [
- "bindgen_cuda",
-]
-
-[[package]]
-name = "candle-nn"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3045fa9e7aef8567d209a27d56b692f60b96f4d0569f4c3011f8ca6715c65e03"
-dependencies = [
- "candle-core",
- "half",
- "libc",
- "num-traits",
- "rayon",
- "safetensors 0.7.0",
- "serde",
- "thiserror 2.0.18",
-]
-
-[[package]]
-name = "candle-ug"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c22d62be69068bf58987a45f690612739d8d2ea1bf508c1b87dc6815a019575d"
-dependencies = [
- "ug",
- "ug-cuda",
-]
-
 [[package]]
 name = "cassowary"
 version = "0.3.0"
@@ -668,23 +596,12 @@ dependencies = [
  "hybrid-array",
 ]
 
-[[package]]
-name = "cudarc"
-version = "0.17.8"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0bf99ab37ee7072d64d906aa2dada9a3422f1d975cdf8c8055a573bc84897ed8"
-dependencies = [
- "half",
- "libloading 0.8.9",
-]
-
 [[package]]
 name = "cudarc"
 version = "0.19.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f071cd6a7b5d51607df76aa2d426aaabc7a74bc6bdb885b8afa63a880572ad9b"
 dependencies = [
- "float8 0.7.0",
  "half",
  "libloading 0.9.0",
 ]
@@ -871,22 +788,6 @@ dependencies = [
  "syn",
 ]
 
-[[package]]
-name = "dyn-stack"
-version = "0.13.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1c4713e43e2886ba72b8271aa66c93d722116acf7a75555cce11dcde84388fe8"
-dependencies = [
- "bytemuck",
- "dyn-stack-macros",
-]
-
-[[package]]
-name = "dyn-stack-macros"
-version = "0.1.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e1d926b4d407d372f141f93bb444696142c29d32962ccbd3531117cf3aa0bfa9"
-
 [[package]]
 name = "either"
 version = "1.15.0"
@@ -908,18 +809,6 @@ dependencies = [
  "cfg-if",
 ]
 
-[[package]]
-name = "enum-as-inner"
-version = "0.6.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a1e6a265c649f3f5979b601d26f1d05ada116434c87741c9493cb56218f76cbc"
-dependencies = [
- "heck",
- "proc-macro2",
- "quote",
- "syn",
-]
-
 [[package]]
 name = "enumflags2"
 version = "0.7.12"
@@ -1031,28 +920,6 @@ dependencies = [
  "miniz_oxide",
 ]
 
-[[package]]
-name = "float8"
-version = "0.6.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "719a903cc23e4a89e87962c2a80fdb45cdaad0983a89bd150bb57b4c8571a7d5"
-dependencies = [
- "cudarc 0.19.4",
- "half",
- "num-traits",
- "rand 0.9.2",
- "rand_distr 0.5.1",
-]
-
-[[package]]
-name = "float8"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c2d1f04709a8ac06e8e8042875a3c466cc4832d3c1a18dbcb9dba3c6e83046bc"
-dependencies = [
- "half",
-]
-
 [[package]]
 name = "fnv"
 version = "1.0.7"
@@ -1183,244 +1050,6 @@ dependencies = [
  "slab",
 ]
 
-[[package]]
-name = "gemm"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ab96b703d31950f1aeddded248bc95543c9efc7ac9c4a21fda8703a83ee35451"
-dependencies = [
- "dyn-stack",
- "gemm-c32 0.18.2",
- "gemm-c64 0.18.2",
- "gemm-common 0.18.2",
- "gemm-f16 0.18.2",
- "gemm-f32 0.18.2",
- "gemm-f64 0.18.2",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aa0673db364b12263d103b68337a68fbecc541d6f6b61ba72fe438654709eacb"
-dependencies = [
- "dyn-stack",
- "gemm-c32 0.19.0",
- "gemm-c64 0.19.0",
- "gemm-common 0.19.0",
- "gemm-f16 0.19.0",
- "gemm-f32 0.19.0",
- "gemm-f64 0.19.0",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-c32"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f6db9fd9f40421d00eea9dd0770045a5603b8d684654816637732463f4073847"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.18.2",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-c32"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "086936dbdcb99e37aad81d320f98f670e53c1e55a98bee70573e83f95beb128c"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.19.0",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-c64"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dfcad8a3d35a43758330b635d02edad980c1e143dc2f21e6fd25f9e4eada8edf"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.18.2",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-c64"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "20c8aeeeec425959bda4d9827664029ba1501a90a0d1e6228e48bef741db3a3f"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.19.0",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-common"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a352d4a69cbe938b9e2a9cb7a3a63b7e72f9349174a2752a558a8a563510d0f3"
-dependencies = [
- "bytemuck",
- "dyn-stack",
- "half",
- "libm",
- "num-complex",
- "num-traits",
- "once_cell",
- "paste",
- "pulp 0.21.5",
- "raw-cpuid",
- "rayon",
- "seq-macro",
- "sysctl",
-]
-
-[[package]]
-name = "gemm-common"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "88027625910cc9b1085aaaa1c4bc46bb3a36aad323452b33c25b5e4e7c8e2a3e"
-dependencies = [
- "bytemuck",
- "dyn-stack",
- "half",
- "libm",
- "num-complex",
- "num-traits",
- "once_cell",
- "paste",
- "pulp 0.22.2",
- "raw-cpuid",
- "rayon",
- "seq-macro",
- "sysctl",
-]
-
-[[package]]
-name = "gemm-f16"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cff95ae3259432f3c3410eaa919033cd03791d81cebd18018393dc147952e109"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.18.2",
- "gemm-f32 0.18.2",
- "half",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "rayon",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-f16"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3df7a55202e6cd6739d82ae3399c8e0c7e1402859b30e4cb780e61525d9486e"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.19.0",
- "gemm-f32 0.19.0",
- "half",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "rayon",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-f32"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bc8d3d4385393304f407392f754cd2dc4b315d05063f62cf09f47b58de276864"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.18.2",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-f32"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "02e0b8c9da1fbec6e3e3ab2ce6bc259ef18eb5f6f0d3e4edf54b75f9fd41a81c"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.19.0",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-f64"
-version = "0.18.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35b2a4f76ce4b8b16eadc11ccf2e083252d8237c1b589558a49b0183545015bd"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.18.2",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
-[[package]]
-name = "gemm-f64"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "056131e8f2a521bfab322f804ccd652520c79700d81209e9d9275bbdecaadc6a"
-dependencies = [
- "dyn-stack",
- "gemm-common 0.19.0",
- "num-complex",
- "num-traits",
- "paste",
- "raw-cpuid",
- "seq-macro",
-]
-
 [[package]]
 name = "getopts"
 version = "0.2.24"
@@ -1510,7 +1139,6 @@ version = "2.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
 dependencies = [
- "bytemuck",
  "cfg-if",
  "crunchy",
  "num-traits",
@@ -1589,7 +1217,7 @@ dependencies = [
  "reqwest",
  "serde",
  "serde_json",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "ureq",
  "windows-sys 0.60.2",
@@ -1789,7 +1417,7 @@ dependencies = [
  "displaydoc",
  "potential_utf",
  "utf8_iter",
- "yoke 0.8.2",
+ "yoke",
  "zerofrom",
  "zerovec",
 ]
@@ -1856,7 +1484,7 @@ dependencies = [
  "displaydoc",
  "icu_locale_core",
  "writeable",
- "yoke 0.8.2",
+ "yoke",
  "zerofrom",
  "zerotrie",
  "zerovec",
@@ -2117,7 +1745,7 @@ dependencies = [
  "encoding_rs",
  "enumflags2",
  "llama-cpp-sys-2",
- "thiserror 2.0.18",
+ "thiserror",
  "tracing",
  "tracing-core",
 ]
@@ -2265,16 +1893,6 @@ version = "2.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f8ca58f447f06ed17d5fc4043ce1b10dd205e060fb3ce5b979b8ed8e59ff3f79"
 
-[[package]]
-name = "memmap2"
-version = "0.9.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "714098028fe011992e1c3962653c96b2d578c4b4bce9036e15ff220319b1e0e3"
-dependencies = [
- "libc",
- "stable_deref_trait",
-]
-
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -2410,20 +2028,6 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "num"
-version = "0.4.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "35bd024e8b2ff75562e5f34e7f4905839deb4b22955ef5e73d2fea1b9813cb23"
-dependencies = [
- "num-bigint",
- "num-complex",
- "num-integer",
- "num-iter",
- "num-rational",
- "num-traits",
-]
-
 [[package]]
 name = "num-bigint"
 version = "0.4.6"
@@ -2440,7 +2044,6 @@ version = "0.4.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "73f88a1307638156682bada9d7604135552957b7818057dcef22705b4d509495"
 dependencies = [
- "bytemuck",
  "num-traits",
 ]
 
@@ -2459,17 +2062,6 @@ dependencies = [
  "num-traits",
 ]
 
-[[package]]
-name = "num-iter"
-version = "0.1.45"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1429034a0490724d0075ebb2bc9e875d6503c3cf69e235a8941aa757d83ef5bf"
-dependencies = [
- "autocfg",
- "num-integer",
- "num-traits",
-]
-
 [[package]]
 name = "num-rational"
 version = "0.4.2"
@@ -2813,43 +2405,6 @@ dependencies = [
  "prost",
 ]
 
-[[package]]
-name = "pulp"
-version = "0.21.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "96b86df24f0a7ddd5e4b95c94fc9ed8a98f1ca94d3b01bdce2824097e7835907"
-dependencies = [
- "bytemuck",
- "cfg-if",
- "libm",
- "num-complex",
- "reborrow",
- "version_check",
-]
-
-[[package]]
-name = "pulp"
-version = "0.22.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e205bb30d5b916c55e584c22201771bcf2bad9aabd5d4127f38387140c38632"
-dependencies = [
- "bytemuck",
- "cfg-if",
- "libm",
- "num-complex",
- "paste",
- "pulp-wasm-simd-flag",
- "raw-cpuid",
- "reborrow",
- "version_check",
-]
-
-[[package]]
-name = "pulp-wasm-simd-flag"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "40e24eee682d89fb193496edf918a7f407d30175b2e785fe057e4392dfd182e0"
-
 [[package]]
 name = "quote"
 version = "1.0.45"
@@ -2971,15 +2526,6 @@ dependencies = [
  "unicode-width 0.2.0",
 ]
 
-[[package]]
-name = "raw-cpuid"
-version = "11.6.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "498cd0dc59d73224351ee52a95fee0f1a617a2eae0e7d9d720cc622c73a54186"
-dependencies = [
- "bitflags",
-]
-
 [[package]]
 name = "rawpointer"
 version = "0.2.1"
@@ -3017,12 +2563,6 @@ dependencies = [
  "crossbeam-utils",
 ]
 
-[[package]]
-name = "reborrow"
-version = "0.5.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "03251193000f4bd3b042892be858ee50e8b3719f2b08e5833ac4353724632430"
-
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -3040,7 +2580,7 @@ checksum = "a4e608c6638b9c18977b00b475ac1f28d14e84b27d8d42f70e0bf1e3dec127ac"
 dependencies = [
  "getrandom 0.2.17",
  "libredox",
- "thiserror 2.0.18",
+ "thiserror",
 ]
 
 [[package]]
@@ -3280,16 +2820,6 @@ dependencies = [
  "bytemuck",
 ]
 
-[[package]]
-name = "safetensors"
-version = "0.4.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44560c11236a6130a46ce36c836a62936dc81ebf8c36a37947423571be0e55b6"
-dependencies = [
- "serde",
- "serde_json",
-]
-
 [[package]]
 name = "safetensors"
 version = "0.7.0"
@@ -3354,12 +2884,6 @@ version = "1.0.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d767eb0aabc880b29956c35734170f26ed551a859dbd361d140cdbeca61ab1e2"
 
-[[package]]
-name = "seq-macro"
-version = "0.3.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1bc711410fbe7399f390ca1c3b60ad0f53f80e95c5eb935e52268a0e2cd49acc"
-
 [[package]]
 name = "serde"
 version = "1.0.228"
@@ -3669,20 +3193,6 @@ dependencies = [
  "syn",
 ]
 
-[[package]]
-name = "sysctl"
-version = "0.6.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "01198a2debb237c62b6826ec7081082d951f46dbb64b0e8c7649a452230d1dfc"
-dependencies = [
- "bitflags",
- "byteorder",
- "enum-as-inner",
- "libc",
- "thiserror 1.0.69",
- "walkdir",
-]
-
 [[package]]
 name = "system-configuration"
 version = "0.7.0"
@@ -3723,33 +3233,13 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "thiserror"
-version = "1.0.69"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6aaf5339b578ea85b50e080feb250a3e8ae8cfcdff9a461c9ec2904bc923f52"
-dependencies = [
- "thiserror-impl 1.0.69",
-]
-
 [[package]]
 name = "thiserror"
 version = "2.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4"
 dependencies = [
- "thiserror-impl 2.0.18",
-]
-
-[[package]]
-name = "thiserror-impl"
-version = "1.0.69"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4fee6c4efc90059e10f81e6d42c60a18f76588c3d74cb83a0b242a2b6c7504c1"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "thiserror-impl",
 ]
 
 [[package]]
@@ -3864,7 +3354,7 @@ dependencies = [
  "serde",
  "serde_json",
  "spm_precompiled",
- "thiserror 2.0.18",
+ "thiserror",
  "unicode-normalization-alignments",
  "unicode-segmentation",
  "unicode_categories",
@@ -4094,7 +3584,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "786d480bce6247ab75f005b14ae1624ad978d3029d9113f0a22fa1ac773faeaf"
 dependencies = [
  "crossbeam-channel",
- "thiserror 2.0.18",
+ "thiserror",
  "time",
  "tracing-subscriber",
 ]
@@ -4168,52 +3658,12 @@ version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
 
-[[package]]
-name = "typed-path"
-version = "0.12.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e28f89b80c87b8fb0cf04ab448d5dd0dd0ade2f8891bae878de66a75a28600e"
-
 [[package]]
 name = "typenum"
 version = "1.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb"
 
-[[package]]
-name = "ug"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "76b761acf8af3494640d826a8609e2265e19778fb43306c7f15379c78c9b05b0"
-dependencies = [
- "gemm 0.18.2",
- "half",
- "libloading 0.8.9",
- "memmap2",
- "num",
- "num-traits",
- "num_cpus",
- "rayon",
- "safetensors 0.4.5",
- "serde",
- "thiserror 1.0.69",
- "tracing",
- "yoke 0.7.5",
-]
-
-[[package]]
-name = "ug-cuda"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9f0a1fa748f26166778c33b8498255ebb7c6bffb472bcc0a72839e07ebb1d9b5"
-dependencies = [
- "cudarc 0.17.8",
- "half",
- "serde",
- "thiserror 1.0.69",
- "ug",
-]
-
 [[package]]
 name = "unic-char-property"
 version = "0.9.0"
@@ -4596,7 +4046,7 @@ dependencies = [
  "sha2",
  "statrs",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tracing",
  "tracing-subscriber",
  "weaver-trace",
@@ -4619,7 +4069,7 @@ dependencies = [
  "serde_yaml",
  "similar",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-stream",
  "toml",
@@ -4658,7 +4108,7 @@ dependencies = [
  "sha2",
  "syn",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-stream",
  "tokio-util",
@@ -4683,7 +4133,7 @@ dependencies = [
  "serde_yaml",
  "sha2",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tracing",
  "tracing-subscriber",
@@ -4700,28 +4150,8 @@ dependencies = [
 name = "weaver-embedding"
 version = "0.1.0"
 dependencies = [
- "anyhow",
- "async-trait",
- "candle-core",
- "candle-nn",
- "chrono",
- "hyper-util",
- "llama-cpp-2",
- "prost",
- "prost-types",
- "serde",
- "serde_json",
- "tempfile",
- "thiserror 2.0.18",
- "tokenizers",
- "tokio",
- "tokio-stream",
- "tonic",
- "tonic-build",
- "tower",
- "tracing",
  "weaver-core",
- "weaver-inference",
+ "weaver-spu",
 ]
 
 [[package]]
@@ -4747,7 +4177,7 @@ dependencies = [
  "serde_json",
  "serde_yaml",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "tokio-stream",
  "toml",
@@ -4768,27 +4198,37 @@ name = "weaver-spu"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "async-trait",
  "base64 0.22.1",
  "bytemuck",
  "cc",
+ "chrono",
  "clap",
- "cudarc 0.19.4",
+ "cudarc",
  "encoding_rs",
  "half",
  "hf-hub",
+ "hyper-util",
  "llama-cpp-2",
  "parking_lot",
+ "prost",
+ "prost-types",
  "reqwest",
  "rustc-hash 1.1.0",
- "safetensors 0.7.0",
+ "safetensors",
  "serde",
  "serde_json",
+ "sha2",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tiktoken-rs",
  "tokenizers",
  "tokio",
+ "tokio-stream",
  "toml",
+ "tonic",
+ "tonic-build",
+ "tower",
  "tracing",
  "tracing-subscriber",
  "weaver-core",
@@ -4804,7 +4244,7 @@ dependencies = [
  "serde",
  "serde_json",
  "tempfile",
- "thiserror 2.0.18",
+ "thiserror",
  "tokio",
  "toml",
  "tracing",
@@ -5228,18 +4668,6 @@ version = "0.6.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
 
-[[package]]
-name = "yoke"
-version = "0.7.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "120e6aef9aa629e3d4f52dc8cc43a015c7724194c97dfaf45180d2daf2b77f40"
-dependencies = [
- "serde",
- "stable_deref_trait",
- "yoke-derive 0.7.5",
- "zerofrom",
-]
-
 [[package]]
 name = "yoke"
 version = "0.8.2"
@@ -5247,22 +4675,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "abe8c5fda708d9ca3df187cae8bfb9ceda00dd96231bed36e445a1a48e66f9ca"
 dependencies = [
  "stable_deref_trait",
- "yoke-derive 0.8.2",
+ "yoke-derive",
  "zerofrom",
 ]
 
-[[package]]
-name = "yoke-derive"
-version = "0.7.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2380878cad4ac9aac1e2435f3eb4020e8374b5f13c296cb75b4620ff8e229154"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
- "synstructure",
-]
-
 [[package]]
 name = "yoke-derive"
 version = "0.8.2"
@@ -5329,7 +4745,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0f9152d31db0792fa83f70fb2f83148effb5c1f5b8c7686c3459e361d9bc20bf"
 dependencies = [
  "displaydoc",
- "yoke 0.8.2",
+ "yoke",
  "zerofrom",
 ]
 
@@ -5339,7 +4755,7 @@ version = "0.11.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "90f911cbc359ab6af17377d242225f4d75119aec87ea711a880987b18cd7b239"
 dependencies = [
- "yoke 0.8.2",
+ "yoke",
  "zerofrom",
  "zerovec-derive",
 ]
@@ -5355,18 +4771,6 @@ dependencies = [
  "syn",
 ]
 
-[[package]]
-name = "zip"
-version = "7.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c42e33efc22a0650c311c2ef19115ce232583abbe80850bc8b66509ebef02de0"
-dependencies = [
- "crc32fast",
- "indexmap",
- "memchr",
- "typed-path",
-]
-
 [[package]]
 name = "zmij"
 version = "1.0.21"
diff --git a/crates/weaver-embedding/Cargo.toml b/crates/weaver-embedding/Cargo.toml
index ca200ea3..094b1dbf 100644
--- a/crates/weaver-embedding/Cargo.toml
+++ b/crates/weaver-embedding/Cargo.toml
@@ -4,131 +4,21 @@ version.workspace = true
 edition.workspace = true
 rust-version.workspace = true
 license.workspace = true
-description = "Embedding pipelines (chunking, prompt-prefix routing, pooling, late-chunking, gRPC client, native llama.cpp backend) for WeaverTools."
-
-# Crate skeleton landed in PR #248. Subsequent move PRs in
-# Block A of the unified-SPU sprint pull existing pieces in:
-#
-# - `late_chunk.rs` from `weaver-database::chunking::late` (PR #249).
-# - `gguf_backend.rs` from `weaver-inference::gguf_embed` (this PR).
-# - `grpc_client.rs` from `weaver-database::persephone::embedding` (next).
-# - `pin.rs` from `weaver-interface::embedder_pin` (next).
-# - The `Embedder` trait + the new in-process Rust backend
-#   (`embedder_client.rs`) implementing it (Phase 1).
+description = "Deprecated re-export shell — content folded into weaver-spu (encoder side) + weaver-core (Embedder trait) per PR-0.5.B + PR-0.5.D; removed in PR-0.5.E."
 
 [features]
+# Forward feature flags to weaver-spu so consumers that opt into
+# `gguf` (legacy GGUF encoder backend + Persephone gRPC client) via
+# this crate still get the corresponding weaver-spu feature enabled.
+# Removed in PR-0.5.E.
 default = []
-# Native llama.cpp GGUF embedding backend. Behind a feature flag
-# so the gRPC-only consumer path (Python embedder during the
-# migration window) doesn't pull `llama-cpp-2` and its CUDA
-# build dependencies. Mirrors the same-named feature in
-# `weaver-inference` — both crates have an independent `gguf`
-# feature for their respective decoder / encoder use of llama.cpp.
-gguf = ["dep:llama-cpp-2"]
-
-# candle-based encoder backend (Block C of the unified-SPU
-# sprint, per `embedder-oxidization-Spec.md` §6). Pulls candle
-# for safetensors loading + FP16 forward pass and the HF
-# `tokenizers` crate. CPU-only by default; the operator opts into
-# CUDA via the `candle-cuda` feature on top. Gated separately
-# from `gguf` so the migration-window builds (gRPC-only
-# consumers) don't pull candle.
-candle = ["dep:candle-core", "dep:candle-nn", "dep:tokenizers"]
-# CUDA acceleration for the candle backend. Implies `candle`.
-# Pulls `nvcc` + `cudnn` build deps from candle-{core,nn}; the
-# default `candle` feature stays CPU so workspace `cargo check`
-# without CUDA toolchain still passes.
-candle-cuda = ["candle", "candle-core/cuda", "candle-nn/cuda"]
+gguf = ["weaver-spu/gguf"]
 
 [dependencies]
-# `weaver-core` carries the `Embedder` trait + associated types
-# post-PR-0.5.B. `grpc_client::EmbeddingClient`'s `Embedder` impl
-# is in this crate (allowed by orphan rules: trait in weaver-core,
-# concrete type in this crate). `embedder.rs` re-exports the trait
-# during the transition window so existing consumers compile
-# unchanged.
+# Sole runtime deps: weaver-core (for the relocated `Embedder` trait
+# + types) and weaver-spu (for the relocated late_chunk / pin /
+# legacy gRPC + GGUF backends + proto module).
+# `default-features = false` on weaver-spu to avoid feature leakage —
+# this shell crate forwards features explicitly via the `gguf` flag.
 weaver-core = { workspace = true }
-
-# `tracing` and `anyhow` are unconditional — `late_chunk` is a
-# pure-stdlib algorithm but we want module-level instrumentation
-# hooks and ergonomic error returns available for any future
-# addition without re-gating. The `gguf_backend` (feature-gated)
-# uses `anyhow::Result` for its public surface today; future
-# native-backend code in this crate will share the same error
-# convention.
-tracing = { workspace = true }
-anyhow = { workspace = true }
-thiserror = { workspace = true }
-
-# `async-trait` for the [`Embedder`] trait's `async fn` methods.
-# Object-safe-via-`#[async_trait]` so consumers can hold
-# `Arc<dyn Embedder>`.
-async-trait = { workspace = true }
-
-# `pin` (the cohort-identity lock-file module) needs serde for
-# the `EmbedderPin` struct, serde_json for the JSON read/write
-# path, and chrono for stamping `pinned_at` rfc3339.
-serde = { workspace = true }
-serde_json = { workspace = true }
-chrono = { workspace = true }
-
-# `grpc_client` (the Persephone embedding gRPC client) — same
-# transport stack `weaver-database`'s removed `persephone`
-# module used. The Unix-socket transport uses `hyper-util` +
-# `tower::service_fn` to wrap a UDS connection into a tonic
-# `Channel`. `prost` carries the generated proto messages.
-# `tokio` (UnixStream) handles the actual UDS connect.
-tonic = { workspace = true }
-prost = { workspace = true }
-prost-types = { workspace = true }
-hyper-util = { workspace = true }
-tower = { workspace = true }
-tokio = { workspace = true }
-
-# Native llama.cpp backend. Same version + features as
-# `weaver-inference`'s `llama-cpp-2` declaration; both crates
-# load `llama-cpp-2` independently but the singleton `LlamaBackend`
-# they share is still process-global (per
-# `embedder-oxidization-Spec.md` §4.4).
-llama-cpp-2 = { version = "0.1.143", optional = true, default-features = false, features = ["cuda", "dynamic-link", "mtmd"] }
-
-# candle FP16 safetensors backend (gated behind the `candle`
-# feature). `candle-core` carries Tensor / Device / DType;
-# `candle-nn` carries the layer primitives (linear, layer-norm,
-# embedding) used by the XLM-RoBERTa encoder + LoRA adapter
-# wiring per `embedder-oxidization-Spec.md` §6.2 / §6.3. The
-# `candle-cuda` feature on this crate enables `cuda` on both
-# candle crates simultaneously; default `candle` stays CPU so a
-# bare `cargo check --features candle` works without nvcc.
-candle-core = { workspace = true, optional = true }
-candle-nn = { workspace = true, optional = true }
-
-# HF `tokenizers` crate — production reads the same
-# `tokenizer.json` the Python reference uses (Spec §7.1).
-tokenizers = { workspace = true, optional = true }
-
-[build-dependencies]
-# Generates `crate::proto::embedding::*` from the embedding.proto
-# at workspace `proto/persephone/embedding/embedding.proto`. See
-# build.rs for the rationale on duplicate generation alongside
-# `weaver-database`'s build.rs during the migration window.
-tonic-build = { workspace = true }
-
-[dev-dependencies]
-# Tests in `gguf_backend.rs` load a real GGUF via
-# `weaver-inference::gguf::{init_backend, load_model,
-# GgufModelParams}` — those primitives belong to the model-runtime
-# layer (`weaver-inference`), not the embedding-pipeline layer
-# (this crate). The dev-dep keeps tests passing in their new home.
-# Phase 1's `EmbedderClient` impl will make this a regular
-# dependency.
-weaver-inference = { workspace = true, features = ["gguf"] }
-
-# `pin` tests use `tempfile::TempDir` for hermetic file-system
-# fixtures.
-tempfile = { workspace = true }
-
-# `grpc_client` integration tests stand up a mock tonic server
-# over a `TcpListener`-derived stream. `tokio-stream` provides
-# the wrapper.
-tokio-stream = { workspace = true }
+weaver-spu = { workspace = true, default-features = false }
diff --git a/crates/weaver-embedding/build.rs b/crates/weaver-embedding/build.rs
deleted file mode 100644
index cd2b062c..00000000
--- a/crates/weaver-embedding/build.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-//! Compile the Persephone embedding service's `.proto` definitions
-//! for the `weaver-embedding` crate.
-//!
-//! Per `embedder-oxidization-Spec.md` §2-§5, embedding-service types
-//! belong in `weaver-embedding`, not `weaver-database`. This
-//! build script compiles the same `embedding.proto` source that
-//! `weaver-database`'s build.rs also compiles — both crates generate
-//! independent Rust types from the same wire format. Wire-compatible
-//! by construction; the duplication is intentional during the
-//! migration window so external proto consumers in
-//! `weaver-database`'s test suite (e.g. `tests/proto_types.rs`)
-//! continue to work without churn.
-//!
-//! A future cleanup PR (post-Phase-1) removes the embedding proto
-//! generation from `weaver-database` entirely once all consumers
-//! migrate to `weaver_embedding::proto::embedding::*`.
-
-use std::path::PathBuf;
-
-fn main() -> Result<(), Box<dyn std::error::Error>> {
-    let manifest_dir = std::env::var("CARGO_MANIFEST_DIR").expect("CARGO_MANIFEST_DIR not set");
-    let proto_root = PathBuf::from(manifest_dir)
-        .join("../../proto")
-        .canonicalize()
-        .expect("proto/ directory not found — expected at workspace root");
-
-    let protos: Vec<PathBuf> = [
-        // Embedding only. weaver-database keeps generating common /
-        // extraction / training for its own consumers.
-        "persephone/embedding/embedding.proto",
-    ]
-    .iter()
-    .map(|p| proto_root.join(p))
-    .collect();
-
-    tonic_build::configure()
-        .build_server(true)
-        .build_client(true)
-        .compile_protos(&protos, &[&proto_root])?;
-
-    println!("cargo:rerun-if-changed={}", proto_root.display());
-
-    Ok(())
-}
diff --git a/crates/weaver-embedding/src/embedder.rs b/crates/weaver-embedding/src/embedder.rs
deleted file mode 100644
index 926e0f0d..00000000
--- a/crates/weaver-embedding/src/embedder.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! Deprecated re-export — the `Embedder` trait + types relocated to
-//! `weaver-core::embedder` per `docs/specs/weaver-spu-Spec.md` PR-0.5.B.
-//!
-//! New code should import from `weaver_core::embedder::*` directly.
-//! This module remains during the transition window so existing
-//! consumers compile unchanged. Removed in PR-0.5.E along with the
-//! whole `weaver-embedding` crate (folded into `weaver-spu`).
-
-// `Embedder` is a trait — Rust has no stable trait-alias mechanism, so
-// the only available form is `pub use` with a `#[deprecated]` attribute.
-// Caveat: `#[deprecated]` on `pub use` does not reliably emit a warning
-// at consumer call sites (see grpc_client.rs's note on this). Consumers
-// migrating to `weaver_core::embedder::Embedder` should be informed via
-// the spec + commit history rather than relying solely on rustc to
-// flag it. Removed in PR-0.5.E.
-#[deprecated(
-    since = "0.1.0",
-    note = "moved to weaver_core::embedder::Embedder; this re-export goes away in PR-0.5.E"
-)]
-pub use weaver_core::embedder::Embedder;
-
-// Type aliases (not `pub use`) so deprecation warnings fire reliably at
-// consumer call sites. Same underlying type — any consumer that has
-// already migrated to `weaver_core::embedder::EmbedderInfo` keeps
-// working unchanged. Removed in PR-0.5.E.
-#[deprecated(
-    since = "0.1.0",
-    note = "moved to weaver_core::embedder::EmbedderInfo; this alias goes away in PR-0.5.E"
-)]
-pub type EmbedderInfo = weaver_core::embedder::EmbedderInfo;
diff --git a/crates/weaver-embedding/src/lib.rs b/crates/weaver-embedding/src/lib.rs
index 1127a13c..9b80de17 100644
--- a/crates/weaver-embedding/src/lib.rs
+++ b/crates/weaver-embedding/src/lib.rs
@@ -1,107 +1,104 @@
-//! Embedding pipelines for WeaverTools.
+//! Deprecated re-export shell — `weaver-embedding`'s encoder-side
+//! modules folded into `weaver-spu` per `docs/specs/weaver-spu-Spec.md`
+//! PR-0.5.D.
 //!
-//! This crate is the home for everything between the
-//! `weaver-inference` model-runtime layer and the consumers of
-//! embedding vectors (`weaver-database` for substrate writes,
-//! `weaver-core` for the surfacing engine and Notepad). Per the
-//! ontology in issue #166 and `docs/specs/embedder-oxidization-Spec.md`
-//! §2:
+//! New code should import:
+//! - `Embedder` trait + types: `weaver_core::embedder::*`
+//!   (relocated in PR-0.5.B).
+//! - `late_chunk` + `LateChunkConfig` / `LateChunkResult` /
+//!   `late_chunk_embeddings`: `weaver_spu::encoder::*`.
+//! - gRPC client (`EmbeddingClient`, `EmbeddingClientConfig`,
+//!   `EmbeddingEndpoint`): `weaver_spu::encoder::grpc_client_legacy`
+//!   (always available — production embedder backend during the
+//!   migration window; retires in PR-3.A).
+//! - GGUF embedder backend: `weaver_spu::encoder::gguf_backend`
+//!   (gated by `gguf` feature; retires in PR-3.B).
+//! - Cohort-pin guard: `weaver_spu::core::pin` (relocated in
+//!   PR-0.5.D).
+//! - Persephone proto: `weaver_spu::proto::embedding::*` (always
+//!   available; retires in PR-3.A).
 //!
-//! ```text
-//! weaver-embedding ──→ weaver-core, weaver-inference
-//!         (uses LlamaModel + llama.cpp singleton)
-//! ```
-//!
-//! ## Status — skeleton only
-//!
-//! This is the first PR of the Phase 0 refactor (issue #166 / sprint
-//! Block A per `docs/specs/unified-spu-sprint-sequence.md` §4). The
-//! crate exists as a workspace member that builds clean; all
-//! functional code lives in subsequent move PRs that target one
-//! source file at a time.
-//!
-//! Move targets, in dependency order:
-//!
-//! - `late_chunk` ← `weaver-database::chunking::late`
-//! - `gguf_backend` ← `weaver-inference::gguf_embed`
-//! - `grpc_client` ← `weaver-database::persephone::embedding`
-//! - `pin` ← `weaver-interface::embedder_pin`
-//! - `embedder` (the `Embedder` trait + Rust in-process impl) — new code
-//!
-//! Once all moves land and the trait + impl are in place, the
-//! Python embedder service migrates behind a feature flag per
-//! Phase 1 of `embedder-oxidization-Spec.md` §10.
+//! This crate stays in the workspace through the migration window so
+//! existing consumers compile unchanged. Removed in PR-0.5.E along
+//! with `weaver-inference` and the rest of the legacy crates.
 
+// Module-scoped allow — this entire crate is a deprecated shell.
+// Every re-export below points at a stable weaver-spu or weaver-core
+// module. External consumers see deprecation warnings on their own
+// `use` lines.
+#![allow(deprecated)]
 #![doc(html_root_url = "https://docs.rs/weaver-embedding/0.1.0")]
-#![deny(rust_2018_idioms)]
 
-pub mod late_chunk;
+// `Embedder` trait + types (relocated in PR-0.5.B).
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_core::embedder::Embedder; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_core::embedder::Embedder;
 
-/// Persephone gRPC/protobuf definitions — embedding service.
-///
-/// Generated by `build.rs` from the workspace
-/// `proto/persephone/embedding/embedding.proto`. Consumers that
-/// previously imported from `weaver_database::proto::embedding::*`
-/// migrate to `weaver_embedding::proto::embedding::*` post-A.5.
-/// `weaver-database` still generates its own copy of these types
-/// during the migration window for backward-compat with
-/// `weaver-database/tests/proto_types.rs`; a future cleanup PR
-/// removes the duplication once that test migrates too.
-pub mod proto {
-    /// Embedding service — vector embedding generation.
-    pub mod embedding {
-        tonic::include_proto!("persephone.embedding");
-    }
-}
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_core::embedder::EmbedderInfo; this alias goes away in PR-0.5.E"
+)]
+pub type EmbedderInfo = weaver_core::embedder::EmbedderInfo;
 
-/// Typed gRPC client for the Persephone embedding service over
-/// a Unix domain socket or TCP. Moved from
-/// `weaver-database::persephone::embedding`. Phase 1's
-/// `EmbedderClient` (the in-process Rust backend) joins this
-/// module behind a shared `Embedder` trait per spec §3.
-pub mod grpc_client;
+// Late-chunking surface (relocated in PR-0.5.D).
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::late_chunk::LateChunkConfig; this alias goes away in PR-0.5.E"
+)]
+pub type LateChunkConfig = weaver_spu::encoder::LateChunkConfig;
 
-/// The backend-agnostic [`Embedder`] trait — implemented by the
-/// gRPC client today and by the Rust in-process backend post-Phase 1.
-/// Consumers (surfacing engine, Notepad, Pen, sleep stage A,
-/// preseed materialization) hold `Arc<dyn Embedder>` and stay
-/// backend-agnostic.
-pub mod embedder;
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::late_chunk::LateChunkResult; this alias goes away in PR-0.5.E"
+)]
+pub type LateChunkResult = weaver_spu::encoder::LateChunkResult;
 
-// `Embedder` and `EmbedderInfo` re-exports — the trait + types relocated
-// to `weaver-core::embedder` per `docs/specs/weaver-spu-Spec.md` PR-0.5.B.
-// New code should import from `weaver_core::embedder::*`. These re-exports
-// keep existing consumers compiling during the migration; removed in
-// PR-0.5.E with the rest of this crate.
-//
-// `#[allow(deprecated)]` is required here because `embedder::Embedder`
-// and `embedder::EmbedderInfo` are themselves marked deprecated in the
-// `embedder` module — re-exporting them at the crate root would trigger
-// `-D warnings` without this. Narrower allow than a crate-level one;
-// scoped to just this re-export line per the reviewer's "narrower
-// allow" guidance.
-#[allow(deprecated)]
-pub use embedder::{Embedder, EmbedderInfo};
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::late_chunk::late_chunk_embeddings; consumers should call the function via its new path"
+)]
+pub use weaver_spu::encoder::late_chunk_embeddings;
 
-// Native llama.cpp GGUF embedding backend. Behind the `gguf`
-// cargo feature so consumers that only need the gRPC client
-// path (Python embedder migration window) don't pull in
-// `llama-cpp-2`. Phase 1's `EmbedderClient` impl wraps this
-// module's `embed_pooled` against a `LlamaContext` constructed
-// per `embedder-oxidization-Spec.md` §6.
-#[cfg(feature = "gguf")]
-pub mod gguf_backend;
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::late_chunk; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_spu::encoder::late_chunk;
+
+// Cohort-pin guard (relocated in PR-0.5.D).
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::core::pin; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_spu::core::pin;
+
+// gRPC client + proto (always-available; the gRPC client is the
+// production embedder backend during the migration window).
+// Retires in PR-3.A.
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::grpc_client_legacy; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_spu::encoder::grpc_client_legacy as grpc_client;
 
-// Cohort-identity pin — SG5-C harness-lifecycle lock file. Refuses
-// to start the daemon if the live embedder identity has drifted
-// from the recorded pin. Lives in `weaver-embedding` so it sits
-// alongside the loader paths it guards (the cohort-pin guard is
-// embedding-substrate concern, not a CLI concern).
-pub mod pin;
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::proto; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_spu::proto;
+
+// GGUF backend (feature-gated; legacy in-process llama.cpp embedder
+// path; retires in PR-3.B).
+#[cfg(feature = "gguf")]
+#[deprecated(
+    since = "0.1.0",
+    note = "moved to weaver_spu::encoder::gguf_backend; this re-export goes away in PR-0.5.E"
+)]
+pub use weaver_spu::encoder::gguf_backend;
 
-// Re-export the late-chunking surface at the crate root so
-// consumers can write `use weaver_embedding::{LateChunkConfig,
-// LateChunkResult, late_chunk_embeddings};` — matches the shape
-// the (now-removed) `weaver_database::chunking` re-exports
-// presented at the previous home.
-pub use late_chunk::{LateChunkConfig, LateChunkResult, late_chunk_embeddings};
+// Convenience top-level re-exports preserving the prior shape:
+// `use weaver_embedding::{LateChunkConfig, LateChunkResult,
+// late_chunk_embeddings};` continues to compile, but each name is
+// individually deprecated above so consumers see warnings.
diff --git a/crates/weaver-spu/Cargo.toml b/crates/weaver-spu/Cargo.toml
index 34a5dcd2..304118cd 100644
--- a/crates/weaver-spu/Cargo.toml
+++ b/crates/weaver-spu/Cargo.toml
@@ -44,6 +44,22 @@ llama-cpp-legacy = ["gguf"]
 # / GPU primitives the legacy decoder backend depends on.
 weaver-core = { workspace = true }
 
+# Encoder-side gRPC stack (folded in from weaver-embedding in
+# PR-0.5.D). Used by `encoder::grpc_client_legacy` to talk to the
+# Python `weaver-embedder.service` during the migration window.
+# Always compiled — `grpc_client_legacy` is the **production
+# embedder backend** today, not a feature-gated path. Retires in
+# PR-3.A alongside the gRPC client.
+tonic = { workspace = true }
+prost = { workspace = true }
+prost-types = { workspace = true }
+hyper-util = { workspace = true }
+tower = { workspace = true }
+async-trait = { workspace = true }
+serde_json = { workspace = true }
+chrono = { workspace = true }
+sha2 = { workspace = true }
+
 # HuggingFace ecosystem (decoder side: download, safetensors, tokenizer)
 hf-hub = "0.4"
 safetensors = { workspace = true }
@@ -70,7 +86,7 @@ reqwest = { workspace = true }
 
 # Serialization
 serde = { workspace = true }
-serde_json = { workspace = true }
+# serde_json declared in the encoder gRPC stack block above.
 
 # Logging
 tracing = { workspace = true }
@@ -97,9 +113,17 @@ tracing-subscriber = { workspace = true }
 # `weaver-inference` had a `cc` build-dep for compiling its custom
 # CUDA kernels via build.rs. Folds in unchanged via PR-0.5.C.
 cc = "1"
+# `weaver-embedding` had a `tonic-build` build-dep for compiling the
+# Persephone embedding-service proto. Folded in via PR-0.5.D; retires
+# alongside the gRPC client in PR-3.A.
+tonic-build = { workspace = true }
 
 [dev-dependencies]
 tempfile = { workspace = true }
+# `tokio-stream` for the legacy grpc_client integration tests folded
+# in from weaver-embedding in PR-0.5.D (the test spins up an
+# in-process tonic server backed by a TcpListener-derived stream).
+tokio-stream = { workspace = true }
 
 [[bin]]
 name = "weaver-infer"
diff --git a/crates/weaver-spu/build.rs b/crates/weaver-spu/build.rs
index f51369bc..da55046c 100644
--- a/crates/weaver-spu/build.rs
+++ b/crates/weaver-spu/build.rs
@@ -1,8 +1,32 @@
-fn main() {
-    // Only compile CUDA kernels when the cuda feature is enabled
+//! `weaver-spu` build script — combines:
+//!
+//! 1. The legacy CUDA-kernel compile step (folded in from
+//!    `weaver-inference` in PR-0.5.C; gated by `cuda` feature, since
+//!    the kernels only get linked when the cudarc-backed decoder
+//!    path is in scope).
+//! 2. The Persephone embedding-service proto compile step (folded
+//!    in from `weaver-embedding` in PR-0.5.D; **always compiled**
+//!    because `encoder::grpc_client_legacy` — the production
+//!    embedder backend during the migration window — depends on
+//!    the generated proto types unconditionally).
+//!
+//! The CUDA step is skipped when `cuda` is not set. The proto step
+//! always runs. Default `cargo build` (no features) does the proto
+//! compile only.
+
+use std::path::PathBuf;
+
+fn main() -> Result<(), Box<dyn std::error::Error>> {
     if std::env::var("CARGO_FEATURE_CUDA").is_ok() {
         compile_cuda_kernels();
     }
+    // Persephone proto is always compiled — `encoder::grpc_client_legacy`
+    // is the production embedder backend during the migration window
+    // (talks to the Python `weaver-embedder.service`), not a
+    // feature-gated path. Retires alongside the gRPC client in
+    // PR-3.A.
+    compile_persephone_proto()?;
+    Ok(())
 }
 
 fn compile_cuda_kernels() {
@@ -38,3 +62,35 @@ fn compile_cuda_kernels() {
     println!("cargo:rustc-link-lib=dylib=cudart");
     println!("cargo:rustc-link-lib=dylib=cublas");
 }
+
+fn compile_persephone_proto() -> Result<(), Box<dyn std::error::Error>> {
+    let manifest_dir = std::env::var("CARGO_MANIFEST_DIR").expect("CARGO_MANIFEST_DIR not set");
+    let proto_root = PathBuf::from(manifest_dir)
+        .join("../../proto")
+        .canonicalize()
+        .expect("proto/ directory not found — expected at workspace root");
+
+    // Embedding-only. `weaver-database` keeps generating common /
+    // extraction / training for its own consumers; this proto step
+    // retires alongside the gRPC client in PR-3.A.
+    let protos: Vec<PathBuf> = ["persephone/embedding/embedding.proto"]
+        .iter()
+        .map(|p| proto_root.join(p))
+        .collect();
+
+    tonic_build::configure()
+        .build_server(true)
+        .build_client(true)
+        .compile_protos(&protos, &[&proto_root])?;
+
+    // Per-proto rerun-if-changed so editing any individual `.proto`
+    // forces a rebuild — `proto_root` alone catches new files added
+    // to the directory but not edits to existing ones reliably
+    // across cargo versions.
+    for proto in &protos {
+        println!("cargo:rerun-if-changed={}", proto.display());
+    }
+    println!("cargo:rerun-if-changed={}", proto_root.display());
+
+    Ok(())
+}
diff --git a/crates/weaver-spu/src/core/mod.rs b/crates/weaver-spu/src/core/mod.rs
index 49d915ea..6b7633ea 100644
--- a/crates/weaver-spu/src/core/mod.rs
+++ b/crates/weaver-spu/src/core/mod.rs
@@ -24,6 +24,7 @@
 
 pub mod gpu_orchestrator;
 pub mod nvlink;
+pub mod pin;
 pub mod probe;
 pub mod vram;
 
diff --git a/crates/weaver-embedding/src/pin.rs b/crates/weaver-spu/src/core/pin.rs
similarity index 100%
rename from crates/weaver-embedding/src/pin.rs
rename to crates/weaver-spu/src/core/pin.rs
diff --git a/crates/weaver-embedding/src/gguf_backend.rs b/crates/weaver-spu/src/encoder/gguf_backend.rs
similarity index 100%
rename from crates/weaver-embedding/src/gguf_backend.rs
rename to crates/weaver-spu/src/encoder/gguf_backend.rs
diff --git a/crates/weaver-embedding/src/grpc_client.rs b/crates/weaver-spu/src/encoder/grpc_client_legacy.rs
similarity index 98%
rename from crates/weaver-embedding/src/grpc_client.rs
rename to crates/weaver-spu/src/encoder/grpc_client_legacy.rs
index 0d2cc722..0ccb6723 100644
--- a/crates/weaver-embedding/src/grpc_client.rs
+++ b/crates/weaver-spu/src/encoder/grpc_client_legacy.rs
@@ -511,11 +511,11 @@ impl std::fmt::Debug for EmbeddingClient {
 // in-process impl post-Phase 1 cutover, with no consumer-side changes.
 //
 // The `info()` impl converts the proto `InfoResponse` to the
-// backend-agnostic [`crate::embedder::EmbedderInfo`] so the trait
+// backend-agnostic [`weaver_core::embedder::EmbedderInfo`] so the trait
 // surface stays clean of proto types.
 
 #[async_trait::async_trait]
-impl crate::embedder::Embedder for EmbeddingClient {
+impl weaver_core::embedder::Embedder for EmbeddingClient {
     async fn embed(
         &self,
         texts: &[String],
@@ -535,9 +535,9 @@ impl crate::embedder::Embedder for EmbeddingClient {
         EmbeddingClient::embed_late_chunked(self, text, task).await
     }
 
-    async fn info(&self) -> Result<crate::embedder::EmbedderInfo, EmbeddingError> {
+    async fn info(&self) -> Result<weaver_core::embedder::EmbedderInfo, EmbeddingError> {
         let resp = EmbeddingClient::info(self).await?;
-        Ok(crate::embedder::EmbedderInfo {
+        Ok(weaver_core::embedder::EmbedderInfo {
             model_name: resp.model_name,
             model_loaded: resp.model_loaded,
             dimension: resp.dimension,
diff --git a/crates/weaver-embedding/src/late_chunk.rs b/crates/weaver-spu/src/encoder/late_chunk.rs
similarity index 100%
rename from crates/weaver-embedding/src/late_chunk.rs
rename to crates/weaver-spu/src/encoder/late_chunk.rs
diff --git a/crates/weaver-spu/src/encoder/mod.rs b/crates/weaver-spu/src/encoder/mod.rs
index db5770e7..303287f2 100644
--- a/crates/weaver-spu/src/encoder/mod.rs
+++ b/crates/weaver-spu/src/encoder/mod.rs
@@ -1,32 +1,59 @@
 //! Encoder half of the SPU — Jina V4 embedder runtime.
 //!
-//! Per `docs/specs/weaver-spu-Spec.md` §2.3 + §3, this module hosts:
+//! Per `docs/specs/weaver-spu-Spec.md` §2.3 + §3. PR-0.5.D folded
+//! in the existing `weaver-embedding` modules:
 //!
-//! - `lora_dispatch` — `MultiAdapterLinear` analogue, a
-//!   PEFT-compatible multi-adapter LoRA dispatcher. Holds 3
-//!   task-keyed `(lora_A, lora_B)` pairs simultaneously and
-//!   dispatches per-call on `task_label`. Phase 1 PR-1.C.
-//!   Companion upstream contribution to `candle-nn` /
-//!   `candle-lora`.
-//! - `pooling` — attention-mask weighted mean + L2 norm. Produces
-//!   the 2048-dim `single_vec_emb` from per-token hidden states.
-//!   Phase 1 PR-1.D.
-//! - `matryoshka` — truncate-and-renormalize for matryoshka
-//!   dimensions `{128, 256, 512, 1024}`. Phase 1 PR-1.D.
-//! - `late_chunk` — late-chunking algorithm. Folds in from
-//!   `weaver-embedding::late_chunk` during PR-0.5.D (algorithm
-//!   unchanged; just relocates).
-//! - `jina_v4` — Jina V4 wrapper that loads base shards + adapter
-//!   shard, attaches `MultiAdapterLinear` wrappers per §3.2,
-//!   implements the 4-string API contract via internal 4→3
-//!   adapter mapping + prefix-prepending. Phase 1 PR-1.D.
-//! - `client` — `EmbedderClient`, the in-process
-//!   `weaver_core::embedder::Embedder` impl backed by candle.
-//!   Phase 1 PR-1.F.
-//! - `grpc_client_legacy` — legacy gRPC client to the Python
-//!   `weaver-embedder.service` (rollback target during the
-//!   migration window). Folds in from `weaver-embedding::grpc_client`
-//!   during PR-0.5.D; deleted in Phase 3 PR-3.A.
+//! - `late_chunk` — late-chunking algorithm (mean-pool token
+//!   embeddings into per-chunk vectors). Algorithm unchanged from
+//!   the prior `weaver-embedding::late_chunk`.
+//! - `grpc_client_legacy` — gRPC client to the Python
+//!   `weaver-embedder.service`. **Production embedder backend
+//!   during the migration window**, always compiled (not
+//!   feature-gated). Renamed during the move (was
+//!   `weaver-embedding::grpc_client`) to make its transitional
+//!   status explicit. Deleted in Phase 3 PR-3.A.
+//! - `gguf_backend` (feature-gated `gguf`) — legacy llama.cpp-based
+//!   in-process GGUF embedder backend (an experimental path; not
+//!   the production embedder). Pulls llama-cpp-2 only when the
+//!   feature is enabled. Deleted in Phase 3 PR-3.B.
 //!
-//! Skeleton only — sub-modules land in their respective Phase 0.5 /
-//! Phase 1 PRs.
+//! ## Planned modules (Phase 1 — not yet present)
+//!
+//! Phase 1 PRs (PR-1.B onwards per
+//! `docs/specs/weaver-spu-Spec.md` §10) add the candle-backed
+//! encoder modules listed below. These are **not importable
+//! today** — they're documented here so the eventual code shape
+//! is visible in the rustdoc, and so that PR reviewers can spot
+//! drift between the planned API surface and what lands.
+//!
+//! - `lora_dispatch` — `MultiAdapterLinear` analogue (planned;
+//!   PR-1.C, companion upstream contribution).
+//! - `pooling` — attention-mask weighted mean + L2 norm (planned;
+//!   PR-1.D).
+//! - `matryoshka` — truncate-and-renormalize (planned; PR-1.D).
+//! - `jina_v4` — Jina V4 wrapper (planned; PR-1.D).
+//! - `client` — `EmbedderClient` (planned; PR-1.F).
+
+pub mod late_chunk;
+
+// `grpc_client_legacy` is NOT feature-gated: it's the production
+// embedder backend during the migration window (talks to the Python
+// `weaver-embedder.service` over Unix socket gRPC). Always-available
+// because the daemon needs an embedder regardless of features.
+// Retires in PR-3.A.
+pub mod grpc_client_legacy;
+
+// `gguf_backend` is feature-gated because it pulls llama-cpp-2 (a
+// substantial transitive build cost). Used only by experimental
+// in-process GGUF embedder paths; not the production path. Retires
+// in PR-3.B.
+#[cfg(feature = "gguf")]
+pub mod gguf_backend;
+
+// Re-export of the late-chunking surface that the prior
+// `weaver_embedding::lib` re-exported at the crate root. Consumers
+// that did `use weaver_embedding::{LateChunkConfig, LateChunkResult,
+// late_chunk_embeddings};` continue to work via the deprecated
+// re-export shell in `weaver-embedding`; this re-export keeps the
+// shape consistent within `weaver-spu` itself.
+pub use late_chunk::{LateChunkConfig, LateChunkResult, late_chunk_embeddings};
diff --git a/crates/weaver-spu/src/lib.rs b/crates/weaver-spu/src/lib.rs
index 706e60fb..8f99bb84 100644
--- a/crates/weaver-spu/src/lib.rs
+++ b/crates/weaver-spu/src/lib.rs
@@ -53,3 +53,18 @@ pub mod core;
 pub mod decoder;
 pub mod encoder;
 pub mod models;
+
+/// Persephone gRPC/protobuf definitions — embedding service.
+///
+/// Generated by `build.rs` from the workspace
+/// `proto/persephone/embedding/embedding.proto`. Folded in from
+/// `weaver-embedding::proto` via PR-0.5.D. Always available (the
+/// gRPC client to the Python embedder is the production backend
+/// during the migration window, not a feature). Removed in Phase 3
+/// PR-3.A alongside the gRPC client itself.
+pub mod proto {
+    /// Embedding service — vector embedding generation.
+    pub mod embedding {
+        tonic::include_proto!("persephone.embedding");
+    }
+}
diff --git a/crates/weaver-embedding/tests/grpc_client.rs b/crates/weaver-spu/tests/grpc_client_legacy.rs
similarity index 96%
rename from crates/weaver-embedding/tests/grpc_client.rs
rename to crates/weaver-spu/tests/grpc_client_legacy.rs
index 969e900d..7d596597 100644
--- a/crates/weaver-embedding/tests/grpc_client.rs
+++ b/crates/weaver-spu/tests/grpc_client_legacy.rs
@@ -8,7 +8,7 @@
 //! surface (Embed, EmbedLateChunked, Info, ensure_ready, GracefulShutdown).
 
 // Module-scoped `allow(deprecated)` — these tests reference the
-// `weaver_embedding::grpc_client::{EmbedResult, EmbeddingError, ...}`
+// `weaver_spu::encoder::grpc_client_legacy::{EmbedResult, EmbeddingError, ...}`
 // type aliases that were marked `#[deprecated]` in PR-0.5.B (they now
 // alias `weaver_core::embedder::*`). The tests stay on the legacy path
 // because they're testing the legacy gRPC client; the whole crate
@@ -18,7 +18,7 @@
 
 use std::path::PathBuf;
 
-use weaver_embedding::grpc_client::{EmbedResult, EmbeddingClientConfig, EmbeddingEndpoint};
+use weaver_spu::encoder::grpc_client_legacy::{EmbedResult, EmbeddingClientConfig, EmbeddingEndpoint};
 
 #[test]
 fn test_config_defaults() {
@@ -58,7 +58,7 @@ fn test_endpoint_variants() {
 
 #[tokio::test]
 async fn test_connect_nonexistent_socket_fails() {
-    use weaver_embedding::grpc_client::EmbeddingClient;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingClient;
 
     let dir = tempfile::tempdir().expect("failed to create tempdir");
     let socket_path = dir.path().join("nonexistent.sock");
@@ -100,10 +100,10 @@ mod mock_server {
     use tokio_stream::wrappers::TcpListenerStream;
     use tonic::{Request, Response, Status, transport::Server};
 
-    use weaver_embedding::proto::embedding::embedding_service_server::{
+    use weaver_spu::proto::embedding::embedding_service_server::{
         EmbeddingService, EmbeddingServiceServer,
     };
-    use weaver_embedding::proto::embedding::{
+    use weaver_spu::proto::embedding::{
         EmbedLateChunkedRequest, EmbedLateChunkedResponse, EmbedRequest, EmbedResponse, Embedding,
         GracefulShutdownRequest, GracefulShutdownResponse, InfoRequest, InfoResponse, LateChunk,
     };
@@ -308,8 +308,8 @@ mod mock_server {
 /// Shared helper: build a client against the mock server's TCP URL. Every
 /// integration test needs the same `EmbeddingClientConfig { endpoint: Tcp(…),
 /// ..Default::default() }` dance, so centralize it here.
-async fn connect_tcp_client(url: String) -> weaver_embedding::grpc_client::EmbeddingClient {
-    use weaver_embedding::grpc_client::{
+async fn connect_tcp_client(url: String) -> weaver_spu::encoder::grpc_client_legacy::EmbeddingClient {
+    use weaver_spu::encoder::grpc_client_legacy::{
         EmbeddingClient, EmbeddingClientConfig, EmbeddingEndpoint,
     };
     EmbeddingClient::connect(EmbeddingClientConfig {
@@ -431,7 +431,7 @@ async fn test_embed_late_chunked_roundtrip() {
 async fn test_embed_late_chunked_rejects_zero_dimension_with_values() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,
@@ -481,7 +481,7 @@ async fn test_ensure_ready_ok_when_loaded() {
 async fn test_ensure_ready_rejects_unloaded_model() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: false,
@@ -550,7 +550,7 @@ async fn test_ensure_ready_maps_unavailable_status_to_not_available() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
     use tonic::Status;
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,
@@ -580,7 +580,7 @@ async fn test_ensure_ready_maps_deadline_exceeded_status_to_not_available() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
     use tonic::Status;
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,
@@ -605,7 +605,7 @@ async fn test_ensure_ready_maps_failed_precondition_status_to_not_available() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
     use tonic::Status;
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,
@@ -634,7 +634,7 @@ async fn test_ensure_ready_propagates_other_status_codes_as_transport() {
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
     use tonic::Status;
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,
@@ -663,7 +663,7 @@ async fn test_ensure_ready_maps_dropped_server_to_not_available() {
     // `ensure_ready`'s match.
     use mock_server::{MockService, spawn};
     use std::sync::{Arc, atomic::AtomicU32};
-    use weaver_embedding::grpc_client::EmbeddingError;
+    use weaver_spu::encoder::grpc_client_legacy::EmbeddingError;
 
     let srv = spawn(MockService {
         model_loaded: true,