fix(validation): prevent validation lead from executing arbitrary local commands

[ganesh47]

Motivation

• The validation lead could return arbitrary shell commands in its LLM-produced plan which were executed by runCommandSet via the system shell, creating a path for command injection and host compromise.
• The intent is to remove the trust boundary that directly executes model-supplied commands while preserving the deterministic, inferred local checks used by validation.

Description

• Add selectApprovedLocalValidationCommands(requested, inferred) which dedupes and returns the intersection of the model-requested commands with the deterministic initialPlan.localValidation.commands.
• Use the approved command list (approvedLocalCommands) instead of validationPlan.localValidation.commands when invoking runCommandSet and when populating the requestedCommands field for skipped runs.
• Preserve existing runCommandSet behavior and tool-wrapper preparation so approved inferred commands run unchanged while any additional LLM-emitted commands are ignored.

Testing

• Ran npm test -- test/validation.test.ts, which completed successfully (1 file passed, 6 tests passed).
• Ran npm test -- test/deliver.test.ts; the validation/deliver scenarios executed and produced logs, but the test run output was large and truncated in the execution environment so a final summary could not be captured here.

---

Codex Task