Add reproducible scenario simulation and Stockpyl comparison

[falltwo]

This PR adds an isolated scenario runner for one SKU and one purchase order. Given the same inputs, version, and seed, it produces the same result. The LLM does not participate in any numerical calculation.

The runner executes a deterministic baseline and compares it with Stockpyl using fixed tolerances. Inputs, computed results, and persisted output are validated. NaN, Infinity, arithmetic overflow, invalid containers, and values that cannot be serialized cause the run to fail, and the failure is still recorded for diagnosis.

Three synthetic fixtures cover normal operation, disruption, and substitution. Stockpyl remains pinned in requirements-scenario.txt; requirements-dev.txt includes it for tests and CI without adding it to production dependencies. Model limits, metric definitions, and the zero-demand fill-rate difference are documented in docs/research/scenario_validation.md.

Testing

• Scenario suite: 87 passed
• Remaining regression suite: 327 passed
• Independent adversarial probes: exit 0
• Combined with the Decision Case branch: 454 passed
• GitHub Actions: passed

Scope

This change covers the reproducible research path only. It does not connect to the existing What-if UI, production database, ERP writes, multi-level BOMs, MCP, or DBOS. The fixtures and simulation results are test evidence, not business forecasts or benefit claims.