feat(platform): verify Spark commit-failure recovery

[zhouning]

Summary

• add a local Spark/Iceberg commit-failure rehearsal with a loopback REST proxy that injects HTTP 503 before provider forwarding
• verify zero snapshot, row, and referenced-file drift after the failed call, then one exact visible commit after an explicit retry
• verify direct MinIO inventory, Gravitino readback, cleanup, fail-closed evidence integrity, platform runtime registration, and CI coverage

Validation

• M3-12 validator: passed with contract 6d8944ab80246dc65891aa81118cb8b73f7ecad699be9a2af5e62d8260c41002 and evidence 39571cdac1e4043bcfc2d03a73b2b12ff925210daf8ae36bc640b8cb14d89401
• M3-12 tests: 23 passed
• M3-10 regression: 17 passed
• platform_truth tests: 11 passed
• required platform suite: 764 passed
• Kustomize, Python compile, shell syntax, JSON/YAML, and diff checks: passed
• rehearsal namespace and its two persistent volumes: absent

Scope

This is local Docker Desktop evidence for a known pre-forward failure and one explicit retry. It does not prove uncertain commit reconciliation, network exactly-once behavior, production object storage, protected identity or TLS, cancel or lineage, Flink, full Spark conformance, or production readiness.