feat(sandbox): implement Phase 1 filesystem probe framework (ProbeFilesystem, runner, tests, docs)

[cto-new]

Summary

Introduce an end-to-end filesystem probe framework to ground sandbox capabilities through concrete experiments, establishing epistemic grounding via observed evidence.

Details

• ProbeFilesystem.java implementing 10 filesystem tests with explicit outcomes and diagnostic detail for both success and failure cases.
• sandbox_probe_runner.py to compile, run, and parse probe output, saving results to knowledge_core/experiments for auditable traceability.
• Tests added at tests/test_sandbox_probe_runner.py to validate core data structures and parsing behavior.
• Documentation added: sandbox_probes/README.md, EXPERIMENTAL_FRAMEWORK.md, PHASE1_COMPLETION_SUMMARY.md, QUICK_START.md to describe usage, structure, and results.
• Enforce formatting via Black and lint checks; all tests pass locally.
• Clear path to Phase 2: network probe and additional probes; MCP integration remains a future step.

Rationale

Grounding sandbox capabilities with concrete experiments reduces epistemic risk and provides verifiable evidence about the agent's operating environment. This aligns with the broader protocol-driven approach and supports iterative expansion to more probes.

Testing

• Phase 1 tests validate compilation, execution, and structured logging of filesystem results.
• Linting and formatting have been applied to satisfy pre-commit/style checks.
• Documentation accompanies code to facilitate review and reuse.

Documentation

• sandbox_probes/README.md: User guide for filesystem probe framework.
• EXPERIMENTAL_FRAMEWORK.md: Framework overview and architectural notes.
• PHASE1_COMPLETION_SUMMARY.md: Phase 1 completion details.
• QUICK_START.md: Quick start workflow for experiments.

Next steps

• Phase 2: design and implement a Phase 2 network probe, with MCP-based on-demand test design and execution workflows (Phase 2b).
• Expand the library of probes (network, services, etc.) following the same pattern and ensuring evidence-grounded conclusions.
• Integrate MCP-based test design with the general agent planning loop.

Warning: Task VM test is not fully passing, using partial setup (3/3 steps working)