fix: stop completed CI task redispatch loops

[numnx]

What changed

• Sync the feature branch with the current main release history.
• Normalize GitHub PR check rollups so newer reruns supersede older cancelled or failed observations for the same workflow/check.
• Select CI repair evidence only when the newest branch-matched run is failed.
• Preserve CI-owned and explicit handoff states through status derivation so completed coding work cannot become pending coding again.
• Suppress merge-attention churn while CI owns a task, separate guardrail handoff deduplication by purpose, and resolve stale guardrail handoffs after a task settles.
• Update canonical and published CI troubleshooting documentation.

Why

A completed coding task could enter an orchestration loop when GitHub returned both an older cancelled check run and a newer successful rerun. The gate treated the historical failure as current, marked the task blocked, and a later derivation pass converted that post-coding blocker back to pending work. That launched the coding provider repeatedly and created merge-attention rows that were opened and resolved in the same watch cycle.

Impact

Latest green CI now advances the existing completed work without another coding invocation. Current CI failures remain owned by CI repair or their human handoff, and the dashboard no longer accumulates stale merge or guardrail interventions for settled tasks.

Validation

• pnpm run ci
• Focused regression suites for PR check normalization, CI run selection, task state derivation, attention coordination/deduplication, and guardrail cleanup
• Live runtime restart against an affected non-test task: invocation and dispatch counts remained stable across multiple watch cycles; the task advanced to completed/merged and its stale handoff resolved

Risk and rollback

Risk is limited to CI status normalization and post-coding state/attention transitions. Roll back this commit to restore the prior behavior; no schema migration is involved.