Add explicit live and replay tractability execution

[vaalessi]

Summary

• adds the file-in/file-out command python -m simulation run --mode live|replay --input <path> --output <path>
• keeps replay output explicitly cached and prevents any live-to-replay fallback
• bridges live mode to an already-configured LABrador managed-agent checkout without deploying or mutating it
• returns stable simulation.execution-error.v1 / CANNOT_COMPLETE reason codes for runtime, deployment, credential, dependency, timeout, provider, and invalid-output failures
• vendors the shared interpretability schema from platform-contracts commit 755499b42ab65d3b01f959b11624dd4e61bdd561 with SHA lock
• records the orchestrator-owned 90-minute node timeout

Verification

• python3 -m unittest simulation.test_module — 16 passed
• python3 schemas/test_schema.py — 45 passed
• replay smoke against examples/input.json — small_molecule_tractable, output_origin=cached_dossier, CACHED_DOSSIER
• vendored interpretability SHA-256: ac7b27908688851b4fc3de5e3d31642a6e9d4422b422f57161f2c9ab42c3d6bb

Demo boundary

Draft only. This PR does not deploy or replace a managed agent and does not change the existing demo checkout. A real live-provider call still requires a separately configured runtime root, existing deployment, credentials, and provider access; that deployment-dependent check is intentionally deferred until after the demo.