Per docs/specs/weaver-spu-Spec.md §10 PR-0.5.D. Mechanical refactor;
no behavior change. Moves weaver-embedding's encoder-side modules
(late_chunk, gRPC client, GGUF backend, cohort-pin guard, proto
generation) under weaver-spu. weaver-embedding becomes a deprecated
re-export shell so existing consumers compile unchanged during the
migration window (removed in PR-0.5.E).

Files moved (git mv preserves history):

  weaver-embedding/src/                      weaver-spu/src/
  ─────────────────────                      ───────────────
  late_chunk.rs                            → encoder/late_chunk.rs
  gguf_backend.rs                          → encoder/gguf_backend.rs (gated `gguf`)
  grpc_client.rs                           → encoder/grpc_client_legacy.rs
                                              (renamed; not feature-gated since
                                               this is the production embedder
                                               backend during the migration
                                               window — talks to Python
                                               weaver-embedder.service)
  pin.rs                                   → core/pin.rs
  build.rs                                 → merged into weaver-spu/build.rs
                                              (compiles persephone proto
                                               unconditionally now since
                                               grpc_client_legacy is always
                                               compiled)
  tests/grpc_client.rs                     → tests/grpc_client_legacy.rs

Removed (already moved in PR-0.5.B):
  embedder.rs (deprecated re-export of weaver-core::embedder; the
              re-exports moved to weaver-embedding/src/lib.rs)

weaver-embedding becomes a re-export shell:
- src/lib.rs: deprecated `pub use weaver_core::embedder::*` (Embedder
  trait + EmbedderInfo) and deprecated `pub use weaver_spu::*`
  (late_chunk, pin, grpc_client, gguf_backend, proto). The
  Embedder trait stays as `pub use` (no stable trait-alias);
  data types use deprecated `pub type` aliases for reliable
  consumer-side warnings (matches PR-0.5.B's pattern).
- Cargo.toml: dropped 8+ transitive deps (tonic, prost, hyper-util,
  tower, async-trait, chrono, sha2, llama-cpp-2, tokenizers,
  tempfile, candle-*, etc.). Sole runtime deps now are
  weaver-core (for the trait) and weaver-spu (with
  default-features = false to forward features explicitly).

weaver-spu Cargo.toml gains:
- tonic, prost, prost-types, hyper-util, tower, async-trait,
  serde_json, chrono, sha2 (encoder gRPC stack)
- tonic-build (build-dep, for proto compile)
- tokio-stream (dev-dep, for grpc_client_legacy tests' tonic
  server fixture)

build.rs merged: combines the legacy CUDA-kernel compile (gated by
`cuda` feature, folded in PR-0.5.C) with the persephone proto
compile (always compiled now, since the gRPC client is the
production embedder backend during the migration window).

Internal imports rewritten:
- crate::embedder::* in grpc_client_legacy (referencing the trait
  + types when those still lived in weaver-embedding) →
  weaver_core::embedder::* (the trait's new home post-PR-0.5.B).

Acceptance criteria from spec §10 PR-0.5.D:
- cargo build --workspace --features gguf,embedder clean ✓
- cargo test --workspace clean — all existing embedder tests pass ✓
  (110 in weaver-embedding's deprecated shell + 12 ensure_ready
   tests in the moved grpc_client_legacy.rs integration test
   file + 588 in weaver-core)
- Existing Python embedder still operational via grpc_client_legacy ✓

Dependencies: PR-0.5.A (skeleton), PR-0.5.B (Embedder trait
relocation), PR-0.5.C (weaver-inference fold).
Companion upstream PR: none.

Five CR findings on PR #277 commit edaa11a, all valid. The code
correctly un-gated `encoder::grpc_client_legacy` and the `proto`
module from the `gguf` feature (since the legacy gRPC client is the
production embedder backend during the migration window, not a
feature-gated path). But several doc / comment sites still described
them as feature-gated. Doc-only fixes:

1. `weaver-embedding/src/lib.rs` lines 14-23: corrected the new-code
   import guide. gRPC client + proto noted as "always available",
   not "gated by gguf".

2. `weaver-spu/build.rs` top-of-file doc: corrected the proto step
   description — "always compiled because grpc_client_legacy depends
   on the generated types unconditionally", not gated.

3. `weaver-spu/build.rs` proto compile: added per-proto
   `cargo:rerun-if-changed` lines (in addition to the proto_root
   one) so editing any individual `.proto` triggers a rebuild
   reliably across cargo versions.

4. `weaver-spu/Cargo.toml` encoder gRPC stack comment: replaced
   "Active only when the `gguf` feature is set" with the correct
   "Always compiled — grpc_client_legacy is the production embedder
   backend today, not a feature-gated path."

5. `weaver-spu/src/encoder/mod.rs` doc comment: reordered to put
   `grpc_client_legacy` first (as the production backend, not
   feature-gated) and clarified `gguf_backend` as the experimental
   in-process path that's the actual feature-gated module.

Validated: cargo check --workspace clean.

Per CR finding on 5e29bcf: the encoder/mod.rs doc listed
`lora_dispatch`, `pooling`, `matryoshka`, `jina_v4`, `client` as
modules without flagging that they don't exist yet. Renamed the
section "Planned modules (Phase 1 — not yet present)" with explicit
"planned" labels per module, plus a sentence noting the docs are
intentionally there so reviewers can spot drift between the planned
surface and what lands during Phase 1 PRs.