Preserve Goal execution across foreign slash commands

[ByBrawe]

Problem

A real OpenCode coexistence canary in opencode-loop exposed an ownership edge case: on OpenCode 1.18.x, a plugin-handled foreign slash command can still materialize as a synthetic user/model turn even when that plugin sets noReply. An active Goal then mistook that bridge message for foreground human steering, cancelled/repinned its executor, and could continue under the foreign command's local agent.

Fix

• add a generic foreign-command steering guard with no Loop-specific dependency
• mark the concrete non-Goal slash-command prompt with a one-time random marker only while an unfinished Goal exists
• consume and strip only markers issued by this plugin instance before the core chat.message steering handler
• unknown/spoofed markers remain ordinary foreground chat
• install the guard immediately above the Goal core so outer safety/task/Plan wrappers still run

Regression coverage

• foreign slash command cannot repin/preempt an active Goal
• real human steering still repins/preempts normally
• spoofed marker cannot bypass steering
• sessions without a Goal remain untouched

A real OpenCode foreign-command ownership canary will be added to this branch before merge. This issue was discovered by Loop PR #140.