[Burntgogi]

Closes #591.

Contract and implementation

• Opt in with POST /apply-template and "return_prompt_ids": true: the existing prompt is accompanied by prompt_ids.
• Absent/false/null preserves the old text-only response and does not tokenize. Wrong types return a named boolean 400.
• One render and one shared private Rendered::ids owner for this route and chat; uses existing guarded/marked tokenization, not a re-tokenization of visible text.
• Media placeholders remain unexpanded in both fields; these are pre-media-expansion ids, not expanded multimodal engine input. Documented in docs/serving.md.
• Based on main 938167b, including serve: DeepSeek's official V4.1 prompt encoding #595's shared request preparation. serve: authorize request heads and bound network image decoding #590 remains owned by the other PC/session; no changes copied or pushed there.

Verification

• FAIL-first real-BPE regression: old implementation omits prompt_ids. After the change the response keeps template <|im_start|>/<|im_end|> as special ids while a user's USER_DEFINED <think> is the seven ordinary byte ids.
• Built the bloomery-serve subprocess prerequisite, then ran the exact gate-serve cargo selection through tools/gate.sh: 522 tests passed (338 library + 184 integration, no ignored/skipped tests).
• Actual loopback bloomery-serve HTTP listener, explicitly MockEngine/MockTokenizer: default/false/null shape, opt-in ids, /detokenize roundtrip including Korean text, count equality against /chat/completions/input_tokens, named wrong-type 400, orderly /shutdown: all observed. Real-BPE isolation is covered by the regression above, not claimed for the mock vocabulary; no GPU inference run.
• cargo fmt --all --check, standard cargo clippy -p bloomery-serve --all-targets, check-comments.sh, check-arch.sh: passed. Clippy reports one pre-existing test-only too_many_arguments warning in unchanged genloop.rs:2510; an extra strict -D warnings invocation therefore failed and was not hidden or relaxed in source.

[github-actions]

Verdict: CLEAN: The opt-in response reuses chat's guarded tokenization and preserves the default response.

Checked

Duplicates: 2 new functions checked; no duplicate owner. Rendered::ids owns the extracted encoding logic. Dryer reported 0 candidate pairs.
Complexity: crapper flagged no functions. The route adds an optional response field; the new method moves the existing two-arm match.
Silent failure: 1 default site followed at api.rs:2436. Only absent/null becomes false; wrong types propagate a named 400.
Tests: static mutation check — replacing guarded encoding at api.rs:2460 with ordinary encoding fails the new exact-id assertion; unconditional ids fail its default-shape assertions. Accepting a string flag at api.rs:2436 fails the changed HTTP test's status assertion.
Gates: no pins, thresholds or cases removed or relaxed. The PR body records FAIL-first for the new regression.
Performance: chat retains one encoding per prompt; the route encodes ids only when requested. No decode-step work changes.
Escape catalog: no C1–C4 patterns matched.

Notes

Static review only; tests were not executed.

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

[Burntgogi]

Read-only coordination/integration check complete at the other session's published head d3c938f: T0 is SUCCESS, review APPROVED, mergeable. I did not edit or push this branch.

In a new detached temporary worktree, #603 (ebcb327, based on main including #595) merged with #590 without conflicts in api.rs or docs/serving.md. The real-BPE guarded-ID regression passed on the combined tree. An actual combined MockEngine HTTP listener with a smoke-only API key returned 401 to an unauthorized request head without waiting for its declared body, then served authorized opt-in prompt IDs with a detokenization roundtrip and orderly shutdown. The temporary merge/worktree was removed after verification. No GPU inference or merger of either PR is claimed. This account has READ permission; landing remains with a maintainer/the owning session.

[juliankang4]

@Burntgogi Thanks for the contribution. The review is done, so I'm merging this.