fix(fork,shell_safety): actionable override hint and blank-error crash

[12britz]

Summary

Two small reliability fixes, each with a regression test:

1. fork — blank-error crash hides the failure (fixes #30)

fork/register_callbacks.py:282:

first_line = str(result.error).strip().splitlines()[0]

When a fork's result carries an empty/whitespace error, splitlines() returns [] and [0] raised IndexError. The blanket except swallowed it, so the fork was marked failed with no error message at all.

Now falls back to the raw error string when there is no first line, and lets the _emit_error actually run.

2. shell_safety — override hint is a guaranteed no-op (fixes #25)

The safety callback only ever runs while yolo_mode is already true — it returns None silently otherwise. Yet all three blocking messages told the agent to /set yolo_mode true, wasting a turn. All three now point only at the setting that actually changes behavior.

The exception path was worse: it declared risk HIGH and suggested raising safety_permission_level to high — the level of the very command that could not be assessed — defeating fail-closed. It now explains the assessment failed and points at manual review instead.

Tests

• tests/test_fork_plugin.py — test_fork_reports_blank_error_without_crashing
• tests/test_shell_safety_override_hint.py — new suite asserting the hint never suggests yolo_mode true and the fault path never raises permission to the fault level
• ruff check + ruff format --check clean
• 38 tests across the touched suites pass