fix(status): treat slash-command and bash user-role lines as synthetic

[BruceWW]

Sessions could pin in Thinking ("◉ Think generating reply") forever after /plugin update (and other pure-local slash commands or !bash invocations). Claude Code flushes 2-3 user-role lines for these — a <local-command-caveat> (isMeta:true), a <command-name> line, and a <local-command-stdout> line — none of which prompt the model, so no assistant reply ever arrives to clear last_user_ts_ms.

is_tool_result_user_msg only filtered tool_result block arrays and let every string-content user line through, so the trailing local- command line opened a thinking window that never closed.

Rename to is_synthetic_user_msg and additionally treat as synthetic:

• entries with isMeta: true
• string content starting with <local-command-stdout|stderr|caveat>,   <command-name>, or <bash-input|stdout|stderr>

Also skip these lines when extracting the session's initial_prompt so the title isn't a <command-name>/plugin... dump.

Add regression tests for the /plugin three-line sequence and the !bash <bash-input>/<bash-stdout> sequence.

[graykode]

lgtm — security clean (single-file parser refactor, no new deps/network/shell). the rename to is_synthetic_user_msg and lifting it to take the whole entry reads right since isMeta lives at the entry level, and keeping the "all blocks are tool_result" check conservative (mixed → still a real prompt) preserves the prior safety. nice catch also gating initial_prompt extraction so titles don't end up as <command-name>/plugin... or a bash stdout dump. the two regression tests exercise the exact 3-line /plugin sequence and the <bash-input>/<bash-stdout> pair from the bug report. merged main locally and cargo test is green (123/123).