server: enable checkpoint reuse for recurrent/hybrid models (qwen3next, Mamba)

[localweights]

Summary

Fixes #1762. Ports the fix from ggml-org/llama.cpp PR for the same root cause.

For Gated DeltaNet, qwen3next, and Mamba architectures, llama_kv_cache_seq_pos_min returns the full sequence length, so the existing apply_checkpoint validity check (cur.pos_min < pos_min_thold) always evaluates false. Consequence: no saved checkpoint is ever considered usable, the slot is force-reset, and every follow-up request re-processes the entire prompt from scratch.

A second small bug compounds it: create_checkpoint requires the slot to hold at least 64 tokens before saving. For recurrent models a short follow-up prompt commonly adds far fewer than 64 new tokens, so the checkpoint that would have rescued the next turn never gets created.

Changes

• apply_checkpoint: extend the lambda's validity check to also accept cur.pos_max <= pos_next. Dense transformer paths still trip the original conditions first, so no behavior change for them.
• create_checkpoint: split the minimum-tokens gate by architecture. Recurrent/hybrid models now create a checkpoint at >= 4 tokens; everything else still requires >= 64.

Validation

Tested on Qwen3.6-27B (qwen3next arch) with --multi-token-prediction --draft-max 3 + --reasoning on:

Before this patch:

• Direct identical-prompt probe: cached_tokens: 0 every call
• Multi-pass synthesis (12-section research report): every section paid full prefill, ~50s per section

After this patch:

• Direct identical-prompt probe: cached_tokens: 1713 / 1718 (99.7% reuse)
• Multi-pass synthesis: 92% prefix-cache reuse on shared-prefix calls

Reference

• Issue: [Bug] qwen3next (Qwen3.6): context checkpoints never restored — full prompt re-processing on every turn #1762
• Upstream port reference: ggml-org/llama.cpp PR #22384

[ikawrakow]

As per this comment in the quoted llama.cpp issue, the "fix" is incorrect.

[FNsi]

I don't really think that that "preserve thinking" related to qwen3next...

Though I tend to believe it's kind of jinja problem, still a problem...

[razlani]

I've been working on the same issue independently. Your PR covers the checkpoint search predicate and threshold fixes - I found a third root cause (partial KV delete nukes checkpoints before apply_checkpoint() runs).

Opened #1976 with the additional fix + a prompt cache load guard. Happy to coordinate if you want to merge efforts.