llama.cpp fixes muse-glimmer bug that swallowed tool calls

llama.cpp contributor ruanslv committed a fix for a subtle bug in the Muse Glimmer chat template that caused tool calls to be silently swallowed into content. Muse Glimmer often answers a user and invokes a tool in a single generation. The template uses <|eom|> (end-of-message) when more messages follow in the same turn, and <|eot|> (end-of-turn) only at the very end. The previous parsing logic assumed the user-facing answer was always last, reading until <|eot|>. But when a tool call follows the answer, there's no <|eot|> before the call—so the parser consumed the tool markup as plain content, and no tool_calls were emitted. The tool never executed.

On a tau2-bench telecom run, this bug affected 43 turns across 19 of 114 tasks—roughly 17% of tasks, showing how common the pattern is. The fix changes the final-message rule to stop at <|eom|> and parse everything after as tool calls. The commit also adds models/templates/muse-glimmer.jinja and four parser tests covering: a plain answer, the <|eom|> junction, markup quoted inside an answer remaining content, and tool markup inside a to=self channel that should stay reasoning. Developers using llama.cpp with Muse Glimmer models should update to prevent silent tool failures.