v2.21.0 — skill→orchestrator→sub-agent wiring hardening

Hardens the skill→orchestrator→sub-agent wiring introduced in v2.20.0. A live dispatch test confirmed the bridge works on the skill-trigger path; the residual defect was that wiring was soft — a degraded monolithic run looked identical to a working sub-agent dispatch. No state-machine, transition, or .mjs changes.

Added

• User-visible mode announcement (W1). SKILL.md Orchestrator Role Assumption now requires a one-line announcement of live mode: sub-agent dispatch engaged vs. degraded single-thread fallback. Silent degradation becomes a visible signal.
• Canonical dispatch example (W1). One literal clarification that prose "Spawn ip-X" means an actual agent-tool call (Agent(subagent_type: "ip-explorer", ...)), not in-thread work.

Fixed

• Install drift between the two sync paths (W2). Updating Local Skill now also cps agents into the skill-bundled ~/.claude/skills/iterative-planner/agents/ (authoritative-by-build), with a diff -rq verify and a Validation-Checklist bullet enforcing parity.

Changed

• SKILL.md condition 2 now states explicitly that agents/orchestrator.md is resolved against the harness-announced skill base directory.

Docs

• Sync --pre-step --help leash-cap text to the actual >= 2 condition (was < 2); refresh README skill badge to v2.21.0.

Full changelog: https://github.com/NikolasMarkou/iterative-planner/blob/main/CHANGELOG.md