feat(verification): independently reproduce and rescore findings

[kusonooyasumi]

Summary

• add opt-in, pre-persistence verification that runs a fresh restricted agent for every candidate finding
• require the verifier to create and execute an independent PoC before confirming dynamic findings and reachable dependency CVEs
• preserve non-exercisable dependency advisories, and perform evidence-based confidence/CVSS review when reproduction fails
• persist verification attempts, evidence, and original/final scoring across Markdown, JSON, SARIF, PDF, and the local viewer
• add --verify-findings for one run, alongside persistent STRIX_VERIFY_FINDINGS configuration

Safety and lifecycle

• use an explicit verifier tool allowlist with no reporting, graph, threat-model, coverage, or MCP mutation tools
• accept verdicts only through the validated terminal tool after a qualifying shell or HTTP replay action
• bound attempts, turns, time, and shared budget usage; keep verifier credentials request-local
• serialize final dedupe/persistence so concurrent verification cannot file duplicate reports
• settle background imports before loading report/agent packages to avoid partially initialized SDK modules

The verifier intentionally reuses the authorized scan sandbox and proxy so it can reproduce the original behavior. A separate disposable verifier container is not part of this draft.

Testing

• uv run pytest (1191 passed)
• Ruff and mypy pass
• isolated pyright passes for the verifier modules
• viewer production build passes
• exact startup import sequence passes in 50 fresh processes
• strix --help exposes --verify-findings