fix: enforce layered trust boundaries for forge and message automation

[atomantic]

Summary

Separate the four GitHub schedules into external intake and trusted maintenance, with live repository authority checks instead of self-only author comparisons. External issue bodies, edits and comments now pass complete-input screening, tool-free API analysis and validated server actions; trusted maintenance independently screens outside discussions and retains withheld activity for retry.

Reuse the same configurable boundary for email, iMessage and Signal: local-only private text providers, no tools/redirects/fallback or autofixer escalation, strict output contracts, and no private identity context. PR review defaults to tool-free static analysis; unsupported execution profiles and legacy queued tasks fail closed.

Make Abuse Guard setup actionable with required screening by default, passive status, explicit install/repair, current pinned classifier dependencies, verified complete token-window coverage, and per-source provider/model/limit controls. Preserve recognized prompt upgrades and custom prompts across installs.

Closes #6255
Closes #6256
Closes #6257
Closes #6258

Test plan

• Full server suite: 39,300 passed, 48 skipped; dedicated test-database suites remain separately gated.
• Full client suite: 10,362 passed, 6 skipped; production client build passed.
• Final schedule picker correction: 51 focused tests and lint passed; local review returned a clean verdict.
• Focused workflow tests for authority revocation, external activity, stale snapshots, per-record withholding, strict source IDs, no fallback/redirects/tools, and actionable setup failures.
• Offline Python classifier contract exercised with synthetic dependencies; actual gated model download/inference remains the explicit install canary.
• Reviewed diff and published text for private data; no live settings, messages, model installation or inference changed during development.