Fewer false failures at startup and in doctor

• AI coworkers on Slack keep starting. When Slack refuses to read a configured channel, such as one the app was never invited to, the coworker now starts anyway. Only the refused read is skipped, and a backfill_read_failed warning names it.
• Earlier reads still replay. Content read before a refused channel or thread is still replayed, so a single refusal no longer costs you the rest of the backfill.
• Real faults still surface. Other failures reading channel history or thread replies, including an invalid cursor or a network fault, still fail startup, so the next start retries that gap.
• Restore a skipped channel. Invite the app to the channel to hear it again.
• Doctor respects intentional policies. A brain with only some tools allowed, such as a private brain the coworker can read but not write, now produces a warning that names the withheld tools and exits 0. Only a brain with no allowed tools fails.
• New tools no longer trigger failures. A tool added by a newer release no longer fails doctor, so you do not need to re-sync the policy with sageox-agent memory add first.

Full notes are on the v0.11.1 release.