server : do not cap slot context to training context (#22140)

[jinweihan-ai]

Summary

Fixes #22140.

server_context silently capped each slot's n_ctx to the model's training context, so any user who extended the context via RoPE scaling (YaRN) — the whole point of models like Qwen3 — effectively had their --ctx-size ignored once the slot was created, even though the KV cache had already been sized for the full n_ctx_seq.

This PR drops the cap and keeps only the warning. llama_context itself already logs "n_ctx_seq (...) > n_ctx_train (...) -- possible training context overflow", so users still see the safety signal.

Before

llama_context: n_ctx_seq     = 4096
llama_kv_cache: size =    2.50 MiB (  4096 cells, ...)
srv    load_model: the slot context (4096) exceeds the training context of the model (2048) - capping
slot   load_model: id  0 | task -1 | new slot, n_ctx = 2048   ← halved

After

llama_context: n_ctx_seq     = 4096
llama_kv_cache: size =    2.50 MiB (  4096 cells, ...)
srv    load_model: the slot context (4096) exceeds the training context of the model (2048) - generation quality may degrade beyond the training context unless RoPE scaling is configured
slot   load_model: id  0 | task -1 | new slot, n_ctx = 4096  ← as requested

/props now reports default_generation_settings.n_ctx = 4096 (previously 2048).

Test plan

• Reproduced the bug on master with stories260K.gguf (n_ctx_train = 2048) and -c 4096.
• Verified the patched build preserves the user-requested n_ctx in both the slot init log and the /props endpoint.
• /completion still returns correctly after the change (20 tokens, stop=true, coherent output).

---

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes. This PR was produced in an AI-assisted workflow — an agent helped surface the candidate issue, drafted the patch, and wrote this description; the fix and reproduction were reviewed and verified locally (bug reproduced on master, fix built cleanly, slot n_ctx and /completion checked) before submitting. Human review and validation in the loop.

[ggml-gh-bot]

Hi @jinweihan-ai, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.