[Feature Request] Add fallback prompt for auto-mode classifier soft-denials instead of returning denial to model

Bug Description
When permissions.defaultMode is set to "auto" and the auto-mode classifier soft-denies a tool call, the denial reason is returned to the model, which then attempts to work around the block — often producing subpar results (retrying with degraded approaches, abandoning the task, or silently skipping steps) instead of completing the user's actual intent.

The existing mitigations don't address this well:

• /permissions → r (retry) requires the user to notice the workaround after the fact.
• The 3-consecutive / 20-total denial threshold triggers too late; a single bad workaround can already have been committed.
• Switching to defaultMode: "default" loses all the ergonomic benefits of auto mode.

Proposed: a setting (e.g., permissions.autoModeFallback: "ask") that, on classifier soft-deny, interrupts and surfaces the standard interactive permission prompt to the user instead of returning the denial to the model. This preserves auto mode's low-friction default while ensuring legitimate actions aren't silently degraded when the classifier is wrong

Environment Info

• Platform: linux
• Terminal: WezTerm
• Version: 2.1.118
• Feedback ID: a0cda5af-e0a2-4847-8d28-a14a9a1c3147