skillscript-runtime v0.21.0

The pre-execution contract is now first-class: an agent reads what a skill takes,
returns, requires, and touches before running it, and the human approver sees the
same effect footprint before signing.

• skill_metadata renamed to skill_preflight (clean rename, no alias — pre-
  adoption window). The name teaches the timing: call it BEFORE executing or
  composing. Its response carries the full contract — vars / returns /
  requires + the effectful footprint (connectors, builtins, shell binaries,
  file-write / file-read / unsafe-shell / notify counts) — plus approval-gate
  state and version/lifecycle.
• The same contract mirrors onto every skill_list entry (returns,
  requires, effectful_footprint), so an agent reads a skill's whole I/O +
  effect contract from one discovery call, no per-skill round-trip. Free — the
  source is already parsed to build the entry.
• Dashboard "What this skill touches" approver checklist. The skill detail
  view renders the AST-derived effectful footprint — the same op enumeration the
  capability gate authorizes — right at the approve action, so the operator sees
  the surface they're signing, not a regex guess.
• Secured-mode approval closure moved to the MCP handler (store-agnostic). The
  "skill_write / skill_status can't grant approval without a valid signature"
  guard now fires at the handler ingress, regardless of SkillStore substrate — a
  custom adopter store can no longer persist a forged Approved status. Closes a
  red-team finding (the per-store guard missed custom stores). The execute gate
  always backstopped it; this stops the store from lying.
• Adopter-finding fixes: one shared approval-key resolver (provisioning +
  signing can't diverge); a loud stderr boot-log line when secured mode + passcode
  are set but signing is unwired (reaches headless adopters who never open the
  dashboard); gate-aware Skills-view badges; runtime version shown in the dashboard.
• Workflow-teaching pass across the MCP tool descriptions — each tool teaches
  where it sits in the author loop (learn → discover → draft → commit → approve →
  run → automate → observe) and what comes next.
• Docs overhaul: in-browser approval + footprint checklist in the playbook;
  the approval/secured-mode env-var surface added to the configuration reference;
  retired-v1-token and shipped-strict-filter corrections in the connector-contract
  and sqlite-skill-store references; ARCHITECTURE rewritten to current reality;
  README + index quickstart + default-deny config fixes.