feat(wo): add optional pending-approval gate for work-order writes

[sharanharsoor]

What

approve_workorder, close_workorder, and cancel_workorder currently apply their
target status unconditionally. docs/mcp-servers.md already documents this: "There
is no status-transition guard." This PR adds a separate, opt-in human checkpoint in
front of those three calls, it does not add the transition check itself. The two are
independent and composable.

How it works

• Off by default. Set AOB_REQUIRE_APPROVAL=1 to enable, same convention as the
  existing AOB_READONLY.
• When enabled, the first call to a gated tool returns a pending result with an
  action_id instead of running.
• An operator reviews and approves or denies it with the new scripts/wo_pending.py
  CLI (list / approve / deny), which talks to the same CouchDB-backed pending
  store the server uses.
• The agent must then repeat the identical call. The approval is bound to a SHA-256
  digest of the call's arguments and consumed exactly once, so approving one call
  never covers a different one (different wonum, different reason, etc.).
• If the pending store is unreachable, the gate fails closed (refuses) rather than
  letting the call through.
• Added destructiveHint=True / idempotentHint=False to the three gated tools'
  ToolAnnotations (previously unset on any write tool in this server).

Why

Looked into whether this server has any human-in-the-loop mechanism for its
destructive write tools before building this. It doesn't, AOB_READONLY is the only
related control, and it's a binary on/off that isn't specific to these three
higher-risk transitions. This adds a lightweight, opt-in checkpoint without touching
default behavior for anyone not using the env var.

Testing

• 30/30 WO server tests pass (19 existing + 11 new), including two end-to-end tests
  that call main.approve_workorder() directly and confirm the work order stays
  WAPPR while gated and only transitions after operator approval + an identical
  retry.
• Verified real MCP tool registration (mcp.list_tools()) picks up the new
  annotations correctly, and that AOB_READONLY=1 still behaves as before alongside
  this change.
• ruff format --check clean; no new ruff check findings beyond pre-existing
  typing-style debt already present throughout this file.

Size

This is ~570 lines against the project's stated 300-line preference. Splitting it
would mean reviewing the gate without its tests, or shipping the operator CLI
separately from the feature it operates on, so it's kept as one PR.

[DhavalRepo18]

@sharanharsoor Maybe it's better to create a git issue for any code changes, and we can discuss the reasoning behind certain decisions.