fix: advisory-mode Stop hook emits its finding instead of returning silently

[Nazim22]

Found by dogfooding Leadline on a real Claude Code install.

Bug. In advisory mode the Stop hook returned bare {}. decision.block is false by design in advisory, and the Stop path had no advisory branch (unlike PostToolUse), so it fell through to silence. The finding was still written to the trace, but nothing surfaced live in the session.

Why it mattered. The README tells every new user to start with --dry-run. In exactly that mode the product's headline behavior, that the agent cannot claim "done" while an obligation is unmet, was invisible. A first run showed route corrections and receipt failures but never a completion-gate finding.

Fix. One guarded line, after the existing enforce-block and anti-lockup branches:

if (engine.mode === 'advisory' && decision.message) return { systemMessage: decision.message };

systemMessage rather than hookSpecificOutput.additionalContext: per the Claude Code hook docs, systemMessage is a user-visible warning, while Stop additionalContext continues the agent loop, which would undermine the advisory fail-open invariant.

Verification (independently reproduced by the reviewer):

• npm test 232/232 in a clean worktree
• npm run bench unchanged (62.5% / 61.1% / 0-4 / 1-1 / 0-2)
• Advisory Stop with an unmet obligation now returns {"systemMessage":"Unmet obligation: ... Satisfy it via ... Alternatively, explicitly abstain"} and no decision key, so it still never blocks
• Advisory Stop with nothing outstanding returns {}, so there is no false noise
• Enforce and anti-lockup paths untouched

🤖 Generated with Claude Code