feat(add-orchestrator): install-time divergence detection (--check + overwrite-prompt wiring) (#8)

[vybe]

Closes #8 (add-orchestrator-scoped version; framework-wide follow-up is #12).

(Supersedes PR #11, which was auto-closed when its branch was rebuilt — the original branch had accidentally been stacked on a stray #5 commit from a prior aborted run, pulling sync-fleet-to-head.md into the diff. This branch is rebased clean onto main: exactly the two files below.)

Problem

Re-running /add-orchestrator offered overwrite / skip / cancel per skill with no version comparison, no diff, and no detection that a local copy was customized. Two real failures this week:

• Corbin's /orchestrate carried a hand-written marketing-hierarchy routing block (Eugene directive). A plain overwrite would have destroyed it silently — it had to be re-applied by hand.
• Two skills sat ahead of the bundle for weeks and nobody noticed, because nothing ever compared the two directions (that is issue add-orchestrator: bundle ships stale copies of sync-fleet-to-head (v1.0<1.4) and profile-fleet (v1.4<1.6) — back-port + close the field-hardening back-flow gap #5's whole existence).

What this adds

A read-only --check mode on /add-orchestrator that, per installed runtime skill, reports the three states — and wires the same comparison into Step 4's overwrite prompt so the warning arrives at the moment of decision, not after:

State	Verdict	Meaning
installed < bundled	WARN	upgrade available
installed > bundled	FAIL	back-port candidate — installed copy is ahead of the bundle; overwrite = silent downgrade (the signal #5 went weeks without)
installed == bundled, content differs	WARN	local customization — diff surfaced before overwrite
installed == bundled, identical	PASS	in sync

Reported in the canon-doctor PASS/WARN/FAIL shape with a one-line fleet-readable verdict (orchestrators dispatch it fleet-wide and read only that line). Version compare is numeric-per-component (1.9 < 1.13), tested against all four states end-to-end.

Deliberately left out (minimal scope, per Eugene's ruling)

• The plugin-framework-wide version — every plugin that copies skills into agent repos has this same gap, but that is a larger framework-surface call. Filed as Plugin-framework-wide install-time divergence detection (generalize #8's --check beyond add-orchestrator) #12; this PR is add-orchestrator-only.
• No new on-disk state / skill-registry inventory — the check is stateless: the bundle is the reference, the installed copy is the subject, compared live.
• No cross-version historical diff — the bundle only carries the current template, so a clean local-customization diff is only possible at equal version; the behind case leads with the version gap. Documented as a caveat rather than built around.

Overlap with PR #9 (issue #7) — still OPEN, not yet merged

#9 also bumps this skill and agent-dev/plugin.json. To avoid a version collision I skip 1.21 / 1.14.1 (reserved for #9): add-orchestrator 1.20 -> 1.22, agent-dev plugin 1.14.0 -> 1.14.2. Whichever merges second will hit a mechanical conflict on the SKILL.md frontmatter (changelog prepend) and plugin.json version line — both trivial to reconcile. No template files touched here, so #9's body changes don't overlap.

Files changed: plugins/agent-dev/skills/add-orchestrator/SKILL.md, plugins/agent-dev/.claude-plugin/plugin.json.

🤖 Generated with Claude Code