server router: CLI API key fails for chat when an API-key file is also used #46

Upstream: ggml-org#28820

A router accepts a CLI API key for model listing but rejects it for chat when --api-key-file is also supplied.

Reproduced on spark-a934.local: Linux aarch64, upstream 790cf51aabd61763486050dec7451d9147cb7c61, CPU-only Release build, Qwen2.5-0.5B-Instruct Q4_K_M.

presets.ini:

[*]
model = /home/aman/models/llama-cpp-triage/qwen2.5-0.5b-instruct-q4_k_m.gguf
ngl = 0
ctx-size = 512
parallel = 1
threads = 2
fit = off
load-mode = none
[triage-a]
load-on-startup = true
[triage-b]
load-on-startup = false

keys.txt contains triage-file-one and triage-file-two, one per line.

llama-server --host 127.0.0.1 --port 18913 --offline --no-ui --models-preset presets.ini --models-max 1 --api-key triage-cli --api-key-file keys.txt
curl -i localhost:18913/v1/models -H 'Authorization: Bearer triage-cli'
curl -i localhost:18913/v1/chat/completions -H 'Authorization: Bearer triage-cli' -H 'Content-Type: application/json' -d '{"model":"triage-a","messages":[{"role":"user","content":"Say hi."}],"max_tokens":2,"temperature":0}'

Observed: listing returns 200, chat returns 401 Invalid API Key. Both file keys return 200 for both endpoints. A direct, non-router server accepts all three keys for chat.

Cause: unset_reserved_args() strips LLAMA_API_KEY but preserves LLAMA_ARG_API_KEY_FILE; the spawned child only receives the file source. This filtering was added in 98c1c7a7bf433f1f64dee23f25e7237caab868c9 (ggml-org#18169); runtime introduction has not been bisected.