Add corrupt success case study

[marsloting]

This adds a case study for a failure mode that is easy to miss in agent workflows: corrupt success.

The agent reports done and the producer-side check passes, but the downstream consumer never receives the work because proof was collected upstream of the point where the failure surfaces.

Why this fits this list:

• It is a concrete verification and termination failure, not just a hallucination.
• It connects a production workflow failure pattern to the 2026 "corrupt success" research thread.
• It adds a practical mitigation: evidence for done/written/shipped claims must come from the consumer endpoint, not from an upstream proxy.

Changes:

• Added a new case study: docs/case-studies/agent-corrupt-success-consumer-verification.md
• Added a README entry under Autonomous Agent Failures
• Added the example to docs/failure-modes/verification-termination.md

Validation:

• git diff HEAD~1..HEAD --check
• Scanned changed files for accidental private identity/internal-context terms before publishing