cendor-sdk 1.5.0

V03 Tier-A wiring (plan-guardrails-v03). Additive and backward-compatible. Requires
cendor-guardrails>=1.3.

Added

• task_adherence at the tool_call stage (A3) — the runner now threads the run's originating
  user turn into Context.instruction on the tool-call gate, so a BYO-judge alignment check
  (judge.task_adherence(respond), reused from cendor-guardrails) can compare a proposed tool call
  against the user's intent. Wire it via rules.llm_judge(check, stage="tool_call", action="flag")
  (advisory + fail-open by default; action="block" short-circuits the tool). The judge call rides
  your instrumented client, so its own spend is budgeted + audited. No adherence-rate claim.
• task_adherence re-exported at the SDK top level (from cendor.sdk import task_adherence,
  alongside judge / rules).
• A1/A2 ride along from cendor-guardrails 1.3: rules.spotlight is re-exported through the SDK
  rules surface, and the reserved annotation-parity metadata keys (detected / filtered /
  redacted / severity / …) now land on the SDK's guardrail_decision audit entries with no change
  to the runner or acttrace.