Add opt-in context-window truncation to the chat endpoint

[stikves]

Summary

Adds an opt-in truncation field to /v1/chat/completions so an over-length conversation can be gracefully fit to the model's context window instead of being rejected outright. The default is unchanged (reject with 400), so this is purely additive and eval-safe.

API

One field, three modes:

truncation	Behavior
"off" (default, also "disabled"/"none")	Reject an over-length prompt with 400 — unchanged behavior
"auto"	Drop oldest whole messages (keeping the system message and the newest turn) until the prompt fits maxContextLength
positive integer N (e.g. 500)	As "auto", but target a budget of min(N, maxContextLength) tokens

POST /v1/chat/completions
{ "messages": [ ... ], "truncation": "auto" }   // or "off", or 500

Invalid values (non-positive integer, unknown string) return 400.

Design

• Truncation runs at the message boundary, before the chat template is applied, so it requires no engine or KV-cache changes and cannot affect attention correctness.
• ContextTruncation.apply is a pure policy — the prompt renderer is injected as a closure — so it is unit-tested in isolation with no tokenizer or server state.
• The system message is pinned and the newest turn is always kept; oldest droppable messages go first. A keep-last token trim is the final safety net so an opt-in request never errors, even if the pinned set alone overflows.
• Both the streaming and non-streaming chat paths go through one shared resolvePromptTokens helper.
• Scope: chat only. /v1/completions (the loglikelihood/eval path, which resets per prompt) is deliberately untouched — truncating there would corrupt scoring.

This is the mechanical, lossless-at-the-boundary layer of context-window management. It mirrors the truncate_prompt_tokens request extension found in other local serving engines and the stateless "truncate to fit" convention, without changing the default.

Testing

• ContextTruncationTests — every mode (off fits/overflows, auto no-op/drops-oldest, tokensAt budget cap, safety-net trim).
• ServerAPITruncationTests — decode of the union field (absent → off, string forms, positive integer, rejection of non-positive/unknown).
• Full CoreAILMCommonTests target green (89 tests); swift build --product llm-server clean; swift format clean.