server: fix abort on exact prompt-cache match with hybrid/recurrent models

[kmehan]

Problem

When an incoming prompt exactly matches a slot's cached tokens, the server backs off one token (n_past--) to guarantee at least one token is decoded for logits. The subsequent truncation calls seq_rm with p0 = n_past > 0, which recurrent memory cannot satisfy (the state can't be rewound to a mid-sequence position when the rollback exceeds n_rs_seq), so common_context_seq_rm hits GGML_ABORT and the whole server dies:

common/common.cpp:1472: failed to remove sequence 0 with p0=490, p1=-1

Observed in production with Qwen3.6-35B-A3B (hybrid attention + GatedDeltaNet): any client re-sending an identical prompt with cache_prompt enabled (regenerate, retry, n_predict: 0 scoring) crashes the server.

Reproduction

Any recurrent model, e.g. mamba-130m Q8_0 on CPU:

llama-server -m mamba-130m.q8_0.gguf --port 8199 -c 2048 -np 1
# POST /completion {"prompt": "<~500-token text>", "n_predict": 0, "cache_prompt": true}  — twice, identical

Unpatched, the second request aborts the server:

W slot update_slots: ... need to evaluate at least 1 token for each active slot (n_past = 491, task.n_tokens() = 491)
W slot update_slots: ... n_past was set to 490
common/common.cpp:1472: failed to remove sequence 0 with p0=490, p1=-1

Fix

Include the exact-match case (n_past == slot.prompt.n_tokens()) in the existing checkpoint-restore / full-reprocess branch by relaxing its condition to n_past <= slot.prompt.n_tokens() and extending pos_min_thold by 1 when there are no new tokens, so the upcoming one-token back-off is accounted for. Recurrent/hybrid models now restore a context checkpoint (or fall back to full re-processing) instead of aborting; attention models are unaffected.

This is a faithful backport of upstream ggml-org/llama.cpp ggml-org#23280 (ccee426) and ggml-org#24110 (6f3a9f3); this fork's merge-base predates both.

Verification

Patched, same repro sequence (×3): HTTP 200 every time, server stays alive —

W ... restored context checkpoint (pos_min = 486, pos_max = 486, n_tokens = 487, n_past = 487, ...)

• exact resend: prompt_n=4 (checkpoint restore; falls back to n_past = 0 full reprocess when no checkpoint exists)
• grown (appended) prompt: prompt_n=21 — prefix cache reuse unchanged
• attention-model regression check (stories15M): exact resend prompt_n=1 (unchanged one-token back-off), grown prompt prompt_n=8, no checkpoint restores — no behavior change for non-recurrent models

[TheTom]

Thank you @kmehan, this is a model fix: minimal diff, honest repro, and it backports the exact logic from upstream ggml-org#24110 (merged June 4), so it carries upstream's blessing. Verified the blast radius: full-attention models never reach the inner block (pos_min stays 0 below the threshold), so only SWA/hybrid/recurrent paths change, which is exactly the bug population. Worst case on a checkpoint-less recurrent model is a full reprocess instead of a dead server. Merging.