Let the tool-call opener win over a reasoning close, and treat an absent KV cache as already cleared

[mcharytoniuk]

Two regressions found while upgrading Paddler from 0.12.0 to 0.13.0. Both were bisected against 0.12 on the same machine, model files and prompts.

1. A tool-call opener that is also a reasoning close is swallowed

For Qwen3.5, 0.13 reports reasoning_closes: [[248069], [248058]] while tool_call_open is [248058] — the same token is in both lists, which is correct (<tool_call> really does end reasoning). try_consume_marker_at_tail probed reasoning closes before the tool-call open, so the token was consumed as a reasoning terminator and the tool-call section never opened. Downstream the whole call was classified as content and never reached parse_chat_message.

Measured on the same prompt (283 prompt tokens, greedy sampling, byte-identical model output): 0.12 produced 25 tool-call tokens and a parsed call; 0.13 produced 0 tool-call tokens and 26 content tokens.

Probing the tool-call open first fixes it, and is the better answer inside a reasoning section too: <tool_call> there ends reasoning and starts the call, so landing in ToolCall beats landing in Content. A close marker that is only a reasoning close is unaffected, so the existing suppression of a stray </think> in content still holds.

2. clear_kv_cache fails on models that have no KV cache

nomic-embed-text-v1.5 is non-causal and its context has no memory module, so llama_get_memory returns null. 0.12's clear_kv_cache was infallible and no-opped; 0.13 returns MemoryHandleUnavailable, which broke embedding-only agents.

Mutating an absent KV cache is vacuously satisfied, so clear_kv_cache, clear_kv_cache_seq and kv_cache_seq_keep now succeed with nothing to do. copy_kv_cache_seq still errors, and kv_cache_seq_add / kv_cache_seq_div / kv_cache_seq_pos_max keep their own handling — they ask for or move data rather than assert an end state.

Tests

Written first, confirmed failing, then fixed:

• sampled_token_classifier: two unit tests for a token shared between reasoning_closes and tool_call_open, from both the content and reasoning sections.
• kv_cache_without_memory_module: an LLM test on nomic-embed-text-v1.5 asserting the three mutations succeed.

make clippy, make test.unit (884 passing) and the full llama-cpp-bindings-tests suite (15 model phases, 0 failures) all pass on CUDA.

One thing worth your call:clear_kv_cache and kv_cache_seq_keep can no longer fail, so their Result is now always Ok. Making them infallible would be the honest signature and would match 0.12, but it is a breaking change, so I left them as-is.