Make plan predict deploy, and warn on silently-dropped source

[s0undsystem]

Closes #9, closes #10, closes #12, closes #7. All reported by @jim-t-trevenx in #6.

Plan now predicts deploy (#9, #10)

Invoke-Plan.ps1 was never updated when the deploy engine moved to create-or-leave, so it still described the old update-in-place model. A diff that mispredicts the apply is worse than no diff, since seeing the apply in advance is the entire reason to run plan.

• An AdvancedRule difference is reported as ! RENAME REQUIRED instead of ~ UPDATE, and states that deploy will leave the rule unchanged. The service rejects Set-DlpComplianceRule -AdvancedRule, so an update is not something deploy can perform; promising one sends the operator looking for a fix that never arrives.
• Rules are fetched per policy with -Policy instead of individually by -Identity. That is what makes rules absent from the manifest visible at all, and they are now reported as - REMOVE to match deploy's prune.
• Those removals are DEFERRED under the same condition deploy applies, when a declared rule is missing or in PendingDeletion, since deploy suppresses pruning until every declared rule is in place.
• Rules in PendingDeletion are reported, since deploy skips them.

Compiler warnings (#12, #7)

Two cases where source compiled to less than it appeared to say, with nothing telling the author.

Boolean alert actions. generateAlert: true validates and deploys nothing. The boolean is still accepted, because it is meaningful when codifying an existing tenant rule where a bare true records observed behaviour, but the compiler now says it will not be deployed and how to make it deploy.

Org-wide policies. A policy with no scope.group compiles to org-wide coverage. Scope is deliberately not made a schema requirement, as discussed in #7: org-wide is a legitimate end state and a tool that cannot express it is broken. But it should be a decision rather than the consequence of a missing key, so the compiler names the affected surfaces. Copilot is exempt since it scopes per user.

Warnings go to stderr and never fail the build. unscoped_locations() and boolean_only_actions() are pure and directly tested.

Assistant prompt. It claimed "SCOPE IS REQUIRED", a guarantee the compiler did not provide, which is the inconsistency at the heart of #7. It now tells the model to always scope its own drafts and to ask rather than assume when a request appears to call for org-wide, while describing the compiler's actual behaviour.

The examples were teaching the bug

Turning on the #12 warning immediately flagged the shipped examples. Three policies, the assistant's golden example, and the README quickstart all used generateIncidentReport: true and therefore compiled to nothing, teaching the exact pattern the issue reports. They now use placeholder recipient lists and actually deploy.

example-raw-advanced-rule.yaml deliberately keeps its boolean: actions on a codified rule are observed documentation rather than intent, and the warning correctly does not fire there.

Verification

• 18/18 compiler tests (5 new), 12/12 assistant tests.
• compile.py runs clean with no warnings on the shipped policies, and the manifest now carries GenerateAlert/GenerateIncidentReport where it previously dropped them.
• powershell/*.ps1 pure ASCII and parsing.
• Invoke-Plan.ps1 exercised against mocked S&C cmdlets across five scenarios: in-sync, differing rule, orphan, orphan-with-missing-rule, and PendingDeletion. Confirmed it reports RENAME rather than UPDATE, surfaces orphans, and defers pruning exactly where deploy does.

#13 is left open as an enhancement.