[noahgift]

Supersedes #5411 (fork PR by @guyernest). His commits are kept in this branch's history; this adds the standard on top: fmt, clippy -D warnings on aprender-serve and apr-cli (--lib), and two falsification rows in contracts/apr-serve-openai-compat-v1.yaml (v1.24.0).

• Assistant request messages with content absent or null parse as "" (OpenAI allows it on the tool-call turn); system/user/tool still require it; serialization bytes unchanged.
• max_completion_tokens aliases max_tokens; both together is a serde duplicate field refusal.
• The GGUF router sits behind AuthGate::from_env() (FALSIFY-AUTH-001).

Verified at this head: cargo test -p aprender-serve --lib chat_message_content 5 passed; ... max_completion_tokens 3 passed; cargo test -p apr-cli --lib gguf_router_auth 4 passed; the whole fork diff (4 Rust files) was read first: no build.rs, scripts, workflows, network or secrets.

Refs #5411. #5411 is not closed by this PR; it is closed by hand after this one shows MERGED.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5544 head=99c084ae576a8f2ad5679be3877b75550271d44d verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

NO SEAT, defect at c6d9da8 (aprender-pace, non-author).

Code diff reads fine: ChatMessageWire/TryFrom (assistant-only optional content, other roles still refused), the max_completion_tokens alias, the GGUF router behind the Bearer gate, and the tests. I did not run them.

Defect: contracts/contracts.nt drops 306 triples for 46 contracts whose YAML still exists at this head (e.g. backward-cpu-staging-v1, vram-guard-v1, wgpu-burn-canary-v1, every wgpu-* and vecmat-blis-gemv-v1). The PR's only intended .nt change is the new falsification line. Base e000347 is current main. git diff origin/main <head> -- contracts/contracts.nt shows 306 deletions and 0 for the intended addition. The file shrinks from 35478 to 35172 lines, and backward-cpu-staging-v1 goes from 6 mentions to 0.

Likely a regen run on a tree missing contracts/entrenar/kaizen and other subdirs. Fix: re-run the regen from a full checkout (make readme-sync / the .nt generator), and confirm the diff to the .nt is only the additions for apr-serve-openai-compat-v1 FALSIFY-CHAT-CONTENT-OPTIONAL-5411. I will seat once the .nt is clean.

[noahgift]

seat 1, aprender-w5018 @c6d9da897874759d9d7da69c930bb67d6b90c0d4

APPROVE. Read the full diff at this head and ran the new tests on a detached checkout of it: aprender-serve chat_message_content + max_completion_tokens 8/8 ok; apr-cli gguf_router_auth 4/4 ok.

• ChatMessageWire carries all 5 ChatMessage fields; absent/null content is accepted only for role assistant, other roles still refused; serialization stays the plain derive (byte-identical test).
• alias = "max_completion_tokens": both names present is a serde duplicate-field error, as documented and tested.
• serve_router now goes through build_gguf_router with AuthGate::from_env(); 401/200 behaviour tests cover it. The doc comment lists the GPU/CUDA/APR routers as still ungated; that is fixed separately on fix/5478-gpu-router-authgate (not in this PR).
  No defect found.

🤖 Generated with Claude Code

[noahgift]

seat 1, aprender-w5018 @99c084ae576a8f2ad5679be3877b75550271d44d

Re-seat. Delta c6d9da8..99c084a touches only contracts/contracts.nt: 306 added triple lines, 0 removed, no source or contract-YAML change, so my earlier read and test runs (aprender-serve 8/8, apr-cli gguf_router_auth 4/4 at c6d9da8) still apply. No defect.

🤖 Generated with Claude Code

[noahgift]

seat 1, aprender-pace @99c084ae576a8f2ad5679be3877b75550271d44d

Non-author (0 aprender-pace commits). My earlier defect is fixed: against its merge-base, the head has no contracts/contracts.nt change, so the 306 dropped triples are gone. The head changes only the contract YAML (+2 falsification entries) and four crates files (+402/-11).

Read in full:

• ChatMessageWire plus TryFrom: only an assistant message may omit or null content; system, user and tool are refused with missing field content. A non-string is still a type error, tested for four roles. Serialization is still the plain derive, with a byte-identical test.
• alias = max_completion_tokens: with both names present, serde refuses with duplicate field. This is deterministic and tested.
• build_gguf_router puts the Bearer gate on the real router, /health included. Tests cover no bearer, a wrong bearer, the right bearer reaching the handler (503 No model available, which only a handler returns) and a disabled gate. The doc comment lists the GPU/CUDA/APR starters as still ungated follow-ups, which I accept as stated scope.

APPROVE. I did not run the tests, and the checks showed no FAILURE. Not armed.