[juliankang4]

Part of #59 (plan and PR order in #190), E6: follow-up, /props names the embedding model held beside the chat seat.

What and why

When bloomery-serve holds an embedding model beside the chat model (--embd-model, #323), /props did not say so. A client that asks /props what the server runs saw only the chat seat. This PR adds an embedding_model object with the model's alias and path to the /props response. The field appears only when a side seat is held. The existing chat fields keep their meanings, and the model-list routes are unchanged.

Release notes: Fixed | /props names the embedding model held beside the chat model.

What changes:

• Embed (crates/serve/src/embed.rs) gains a file-path accessor. Model, SideEngine, Stub and MockEmbed implement it.
• The side seat keeps the path from the configuration it already opens (crates/gpu-gates/src/side_seat/gemma_embedding2.rs) and passes it through the existing open-time facts channel (crates/gpu-gates/src/side_seat.rs). Image support passes through unchanged.
• crates/serve/src/embeddings.rs snapshots the path next to the other model facts. crates/serve/src/api.rs adds the optional field to the props response.
• docs/serving.md documents the field.

Why no comparison against a reference engine: the change returns file metadata that the server already holds. It computes no embedding, so there is no vector to compare and no number to time.

Code shape

• No second copy. The path travels through the existing Embed contract and the existing facts channel. No new endpoint, parser, listing renderer or fixture.
• Model code holds only architecture facts. The EmbeddingGemma 2 file only keeps the path it was opened with. The common serve code gained no per-model branch.
• Dead code is gone. Nothing was made unused.
• No silent failure. No new fallback, ignored Result or default. The field is absent only when there is no side seat.
• Comments state what is true now.
• No relaxed gate. No band, tolerance or lint moved. The lint warning count is 41 on main and on this branch.
• cargo fmt and cargo clippy are clean on what I touched: cargo fmt passes, and the full Linux check (85 of 85) adds no lint warning.

Duplication found or removed (path:line), or "none": none.

What I ran

The checks ran on a Linux host, CPU only, on the commit before a patch-identical rebase onto the current upstream/main. The final head ff4493b04 has the same seven changed files as the tested head ea7b96df, and a range-diff shows the one commit as equal. I did not repeat the tests after the rebase.

Check	Result
FAIL-first: the new test api::embeddings::tests::discovery_names_the_seated_models run on the base bdc70824	rc 101. /props has embedding_model as Null, where the test expects {"alias": "mock-embd", "path": "/models/mock-embd.gguf"}
The same test on ea7b96df, with and without a side seat	passes
cargo test --locked -p bloomery-serve --lib on ea7b96df	234 passed; 0 failed
cargo oxide test --arch sm_121 -- -p bloomery-gpu-gates --features deepseek41,gemma-embedding2,vision --release --lib side_seat::tests -- --nocapture	2 passed; 0 failed
Wrong-accessor check: use embed.name() for the path in embeddings::Side::new, then model.name() in SideEngine::spawn, one at a time	each makes its own contract assertion fail (an alias reported as the path), and the restored thread test passes
cargo build --locked -p bloomery-serve --bin bloomery-serve, then the process-level test api::tests::a_key_never_reaches_what_the_process_prints_or_serves	passes
Changed files on the final head against the tested head	all seven are equal, and no tracked file was edited in the trees used

Not run: cargo clippy, contrib-check, and any gate that needs a GPU or a model file. No inference ran.

What I ran into

None of this blocks the PR.

The first full library run failed one existing test (233 passed; 1 failed) because the bloomery-serve binary it needs had not been built. I built it, and the final run above is green. A first run of the side-seat tests in the debug profile failed CUDA code generation in gpu, gpu-vision and gpu-deepseek41, which this PR does not touch. The release profile, which the gate recipes use, compiles and passes.

Hardware (if the change is for a machine)

Not a machine-specific change. The tests ran on a CPU-only Linux host. The oxide test compiles the device-linked crates but runs only host-side stub tests, so no GPU was used and no timing is claimed.

AI help

AI 코딩 에이전트가 코드를 작성한 후 설명을 썼고, 테스트를 마쳤습니다. diff 리뷰와 PR 초안 리뷰는 제가 직접 했습니다.

[github-actions]

Verdict: CLEAN: The embedding path reaches /props through startup metadata, with no inference or model-list behavior change.

Checked

Duplicates: 6 new function or method declarations checked; no duplicate owner. Dryer found 0 pairs in serve; gpu-gates was checked manually.
Complexity: crapper flagged no changed function.
Silent failure: 0 added fallback or ignored-result sites. The path comes from the file opened for the side model.
Tests: static mutation checks cover all 3 added or changed tests. Removing api.rs:1848 fails discovery's side-present assertion; inserting a null field without a side fails its absence assertion. Replacing side_seat.rs:324 with model.name() fails the worker facts assertion. Changing embeddings.rs:137 from [v] to v fails each_route_answers_its_shape. Its removed discovery assertions are retained in the new test.
Gates: no pins or thresholds moved; the PR body reports FAIL-first for discovery.
Escape catalog: no patterns matched.
Adversarial checks: absent side, alias/path substitution, concurrent discovery during inference, and unsupported architecture are blocked by the presence branch, distinct metadata, immutable snapshots, and architecture validation.
Performance: startup and discovery metadata only; no decode, prefill, or kernel work added.

Notes

Tests were not rerun in this read-only review. The PR reports CPU-only validation and supplies no GPU gate evidence.

_{Card check: needs-gpu. This pull request changes crates/gpu-gates/src/side_seat.rs, crates/gpu-gates/src/side_seat/gemma_embedding2.rs and carries no gate run on a card (CONTRIBUTING's evidence list); a maintainer adds gpu-verified after one.}

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}

[github-actions]

Verdict: CLEAN: The embedding model's path reaches /props through startup metadata, with no must-fix defect found.

Checked

Duplicates: 6 new function/method items checked; no duplicated owner. dryer reported 0 candidate pairs.
Complexity: crapper flagged no functions; props gains one branch for the optional side model.
Silent failure: 0 new fallback or ignored-result sites; failed model opens still refuse startup.
Tests: static mutation review covers all three new or changed tests. Removing the field at crates/serve/src/api.rs:1848 fails discovery_names_the_seated_models; using the alias instead of the path also fails its exact object assertion. Substituting model.name() at crates/gpu-gates/src/side_seat.rs:324 fails the thread test's metadata tuple assertion. Replacing the embedding response's model at crates/serve/src/embeddings.rs:161 with the chat alias still fails each_route_answers_its_shape.
Gates: no pins moved or assertions lost; model-list assertions moved into the discovery test. The body reports FAIL-first evidence.
Adversarial checks: absent side, swapped alias/path, concurrent metadata reads, and escaped filenames are covered by optional insertion, distinct metadata, immutable snapshots, and JSON serialization.
Performance: startup and metadata endpoints only; no added inference work.
Escape catalog: no C1–C4 patterns matched.

Notes

dryer skipped crates/gpu-gates; the changed code was checked with Grep.
Tests were not run during this review. Reported runs name ea7b96df; the body's equivalence claim ends at ff4493b04, rather than reviewed head d3e248d0. No GPU gate evidence was supplied.

_{Card check: needs-gpu. This pull request changes crates/gpu-gates/src/side_seat.rs, crates/gpu-gates/src/side_seat/gemma_embedding2.rs and carries no gate run on a card (CONTRIBUTING's evidence list); a maintainer adds gpu-verified after one.}

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}

[juliankang4]

The review is done and the contribution check passed, so I'm merging this.