refactor(skills): extract sections from orchestration/SKILL.md to free Tier-1 budget

[michael-wojcik]

Problem

pact-plugin/skills/orchestration/SKILL.md is currently 610 lines, well past the ~292-line per-file post-compaction restoration budget per memory 444 (4-tier durability model investigation, PR #444 / commit 8621cc7). Content past line ~292 may not restore after compaction, creating eviction risk for critical operating instructions.

This is a hygiene issue independent of any specific feature work — the budget overrun is structural and worsens with every additive edit to the orchestration skill body.

Action

Identify candidate sections to extract into separate skills:

• Agent dispatch templates → skills/pact-agent-dispatch (or expand existing dispatch-related skill)
• Teachback review flow → expand skills/pact-teachback
• Phase-by-phase principles → consolidate elsewhere (own skill or inlined into command files)
• Agent tracking task lifecycle table → could move to a smaller skill or be replaced with a cross-ref

Restore skills/orchestration/SKILL.md to comfortably within the ~292-line budget so that the full operating reference reliably restores after compaction.

Discovered during

#591 plan-mode session 2026-04-29 (plan at docs/plans/inbox-monitor-wake-plan.md). Preparer's R3 risk surfaced the budget overrun while researching the wake-mechanism placement; user explicitly requested this be filed as a separate hygiene issue regardless of #591 outcome.

Acceptance criteria

• skills/orchestration/SKILL.md ≤ 292 lines post-refactor.
• All extracted content preserved in new/existing skill files with no semantic loss.
• No agent-execution-time behavior changes (extracted skills load on the same trigger conditions as before, or via explicit cross-refs from the slimmed orchestration body).
• Test invariant added asserting skills/orchestration/SKILL.md line count ≤ 292 (catches future regression).
• Cross-refs from the slimmed orchestration body to extracted skills follow the slug-link form per memory Canonical Task() dispatch is mirrored inline at every consumer site (PR #457) precedent for cross-surface references.

Related memories

• 444 — 4-tier durability model investigation; defines the ~292-line per-file restoration budget.
• Canonical Task() dispatch is mirrored inline at every consumer site (PR #457) — load-bearing cross-ref discipline.
• feedback_llm_docs_structural_delivery — structural > inline > warn-ref > plain-ref.