UR-Nexus 1.87.0

• Verification now binds command receipts to the actual execution and source
  state. Stale, failed, unrelated or fabricated evidence cannot mark a task
  verified; shell and script mutations participate in change tracking.
• Workflows, crews, background work and persistent goals retain ownership and
  progress across interruption. Goals support model-directed replanning,
  criterion-specific outcomes and optional aggregate budgets without imposing
  a fixed task strategy or a default total invocation ceiling.
• Browser QA executes assertions and interactions in a real browser, records
  viewport screenshots, traces and console errors, and fails on broken behavior.
  Evaluation suites use independent behavioral graders, record provenance and
  false completion, and can invoke real external benchmark runners through
  the Harbor adapter or an explicitly configured runner.
• Memory retrieval searches content with multilingual and structural signals,
  tracks provenance, conflict and expiry, and invalidates stale repository
  entries. Optional embeddings remain optional. Provider profiles record tested
  capabilities, and selective tool discovery works across native providers.
• Added dependency-free Node and Python SDKs with events, approvals, cancellation
  and session resume. Editor integrations use the bidirectional CLI transport;
  remote clients share authenticated ACP sessions and replay cursors.
• Remote sessions persist events and pending approvals, support scoped device
  access and revocation, and expose explicit interrupted states. Optional team
  policy adds roles, shared request quotas, scoped environment secrets, reviewed
  automation and a verifiable local audit trail.
• Added a first-task guide, shared readiness diagnostics and an opt-in sandbox
  preset. Fixed filesystem permission matching and report unknown cloud pricing
  as unknown in SDK results, goal accounting and evaluation reports.
• Source ACP clients tolerate missing build metadata, unattended workspace
  commands avoid unused input pipes, and early process-launch failures return
  explicit errors without replay. Docker HTTP evaluation works with remote
  daemons as well as local engines.
• Updated the embedded runtime to fix descriptor ownership after browser
  subprocesses, which could break unrelated commands in long sessions. Shell
  launch errors now retain the operating system diagnostic.
• Kept one canonical command per capability and updated technical documentation,
  installation and release guidance. Native macOS/Linux, npm and Homebrew
  installations continue to run without a separate Bun installation. Windows
  native publication, Apple notarization and marketplace publication remain
  outside this release pending their respective prerequisites.