Phase 1: Correctness, units, CI, rename, polish

[Gavand1969]

This PR addresses every issue from the code review.

Correctness fixes

• solve_base_stock: removed dead fill_rate line and corrected the Type-II formula. Was 1 - E_B * D / D which is dimensionally wrong; now 1 - E[B per cycle] / mu_L.
• Inventory lead-time units: the previous sigma_L = sigma_d * sqrt(L) mixed years (L) and daily sigma. Standardized everything to days, with lead_time_days in the API. Internal: mu_L = (D/365) * L_days.
• Little's Law check: was raw abs(L - lam*W) which is meaningless at high rho. Now a relative residual |L - lam W| / max(|L|, 1).
• Simulation warm-up: fixed off-by-one (was resetting at last warmup customer, not first measurement customer). Raised default warmup from 10% to max(100, N/5).
• Production CORS: ALLOWED_ORIGIN is now required when ENVIRONMENT=production. Loud failure instead of silent empty list.

Engineering polish

• Cached anthropic.AsyncAnthropic at module level instead of per-request.
• Added pyproject.toml with ruff + mypy + pytest config.
• Added .github/workflows/ci.yml so the test badge in README actually works.

Renames

• Sim2Real -> Sim2Sim everywhere (README, page title, hero, brand, docstrings, CSS, JS).

Tests

• Was 57 tests passing, now 64 passing.
• Added test classes for reorder-point and base-stock covering: basic formulas, monotonicity in service level, zero-variability edge case, fill-rate range invariant.

All tests green locally.