cendor-sdk 1.2.0

Added

• Agent(guardrails=[…]) — a deterministic gate wired at all four stages of the loop. Attach
  cendor.guardrails rules (re-exported: Guardrail, guardrail, GuardrailTripped, rules) to
  gate input (pre-spend), tool_call, tool_output, and output. A block at the input/output
  stage raises GuardrailTripped (fail-closed — an input block refuses before the model is ever
  called, $0 spent); a block at the tool stages returns a "[blocked by <name>] <reason>" tool
  result so the loop continues without the side effect (mirroring require_approval's "[denied]").
  redact rewrites the outgoing messages / tool result / output; flag records and proceeds. Every
  decision emits on the bus, so an attached AuditLog chains it as a guardrail_decision entry —
  correlated with the run's decision, no extra wiring. Works on the sync, async, streaming, and
  multi-agent paths; each agent in a team gates with its own list.
• Per-run override — run(agent, input, guardrails=[…]) (and run.aio) replaces the agent's
  list for that run; guardrails=[] disables gating for the run. For a team, the override applies
  to every segment. This is per-agent/per-run scoped — unlike the process-global guard(), which
  stays the acttrace-policy context manager (the two are distinct).
• Depends on cendor-guardrails>=1.0,<2 (joins the bundled stack).