docs: add agent-harness to skills-extension-candidates

[ar27111994]

Adds agent-harness to the candidates snapshot.

It matches the tracker's inclusion criterion — "ships skills in some non-MCP
form (separate install path, plugin marketplace, framework-native provider)" —
via a dedicated supply-chain CLI (discover → mirror → stage → activate → wire)
across 7 editor hosts.

Its delivery layer already implements the threat model's recommended install
baseline pattern — mirror-then-verify at fetch time, digest-pinned immutable
host-private store — plus an explicit quarantine gate for T1 (prompt injection)
and T2 (host-side execution) before workspace activation, for non-MCP skill
sources today. SEP-2640 support is planned rather than in-flight: client-side
skills/list/skills/get consumption and server-side endpoints are tracked at
ar27111994/agent-harness#492
and #493.

Happy to document the delivery/quarantine implementation as a reference case
for the threat-model corpus if that's useful to the group.