fix(skills): make owner decisions understandable and dependency-closed

[levineam]

Summary

jarvOS now turns a held skill-sync condition into an understandable owner decision instead of exposing a raw machine label such as needs_owner_input. It names the affected skill, explains why sharing stopped, lists the safe choices, and says what happens next. The same decision is available through the owner-gated jarvos_shared_skills interface, and a valid choice is checked against the current source digest before it can change anything.

The sync path also now treats skill dependencies and harness roots as one correctness boundary. A wrapper is not reported as available unless its declared dependency closure is complete, the destination is the harness's enrolled discoverable root, and a fresh observation matches the desired set. Unsafe or unresolved closures preserve existing files and remain non-complete.

Session-settled decisions carried from planning: owner decisions belong in the public skills/runtime-kit contracts, while selected-runtime launcher installation and live activation remain a separate deployment boundary.

Verification

• npm --prefix modules/jarvos-skills test — 157/157 passed.
• npm --prefix modules/jarvos-agent-context test — 64/64 passed, including a valid owner keep-local resolution and stale-reply rejection.
• npm --prefix modules/jarvos-runtime-kit test — passed.
• Isolated four-harness dogfood — Codex, OpenClaw, and Hermes proved model-visible; Claude truthfully remained verification_pending; desired and observed tuples matched; the second run was a no-op.
• Runtime adapter checks — 4/4 passed.
• Package parity — 5/5 passed.
• git diff --check — passed.

Follow-up boundary

The public contract is ready for review. The private selected-runtime branch separately fail-closes owner-decision delivery until the activation session injects the host-bound acknowledgement adapter. No live Telegram delivery is claimed by this PR.

Related: Skill Owner Decisions and Parity Closeout (SUP-3820).