fix(acms): align SkeletonCompressorService.compress() with SkeletonCompressor protocol

- Updated SkeletonCompressorService.compress() to accept
  (fragments: tuple[ContextFragment, ...], skeleton_budget: int)
  -> tuple[ContextFragment, ...], matching the SkeletonCompressor protocol
- Removed skeleton_ratio and CompressionResult from the public API
- Added @runtime_checkable to SkeletonCompressor protocol in acms_service.py
- Added structural subtype assertion at module level to prevent future
  protocol drift
- Rewrote all Behave feature scenarios and step definitions to use
  skeleton_budget
- Updated benchmarks and robot helpers to use absolute token budgets
- Removed CompressionResult export from services __init__.py and
  vulture_whitelist.py
- The depth_breadth_projection.py call site already correctly computed
  and passed skeleton_budget

ISSUES CLOSED: #2925

features/services_init_coverage_r3.feature

@@ -429,11 +429,6 @@ Feature: Services __init__ lazy-import coverage (round 3)
     Then svcov3 the attribute "PersistentSessionService" should be resolved
 
   # ---------- skeleton_compressor ----------
-  Scenario: svcov3 lazy-load CompressionResult
-    Given svcov3 a fresh services module
-    When svcov3 I access lazy attribute "CompressionResult"
-    Then svcov3 the attribute "CompressionResult" should be resolved
-
   Scenario: svcov3 lazy-load SkeletonCompressorService
     Given svcov3 a fresh services module
     When svcov3 I access lazy attribute "SkeletonCompressorService"