feat(investigation): deliver autonomous investigation engine

[FlowerYes]

Summary

• connect knowledge snapshots and bounded plans to the existing distributed runtime and Agent Harness
• collect durable, attempt-bound investigation results and provenance-complete evidence
• evaluate, verify, update knowledge, measure progress, and terminate through explicit stages
• recover ready-wave submissions and resume collected/evaluating/updating sessions without rerunning completed agents
• add SQLite persistence, lifecycle events, scientific metrics, CLI, deterministic 10/50/10 benchmark, ADRs, and architecture docs

Validation

• 325 tests passed
• 87.04% branch coverage
• Ruff format/lint passed
• strict mypy passed for 52 runtime source files
• compileall passed
• deterministic benchmark: 10 gaps, 50 candidates, 10 selected, 10 succeeded distributed tasks