feat: Monte Carlo uncertainty propagation (gpu_stack.uncertainty)

[Cuuper22]

$(cat <<'EOF'

Summary

• Adds gpu_stack/uncertainty.py with a public API for Monte Carlo uncertainty propagation over the existing symbolic resolver
• Three distribution types: uniform(low, high), normal(mean, std), lognormal(mu, sigma) with sign-assumption validation against SymPy variable assumptions
• UncertainAssignment(name, distribution) pairs a registered variable name with a distribution; validated at construction
• propagate_uncertainty(preset_or_assignments, targets, uncertain, n_samples, seed) collects per-target samples through the existing public resolve() path
• Fast path: SymPy lambdify over the symbolic expression (200 samples in <1 ms vs ~14 s per-sample fallback)
• UncertaintyResult / TargetUncertaintyStats result artifacts with to_dict() and echoed input specs
• 35 tests in tests/test_uncertainty.py covering: distribution validation, sign-assumption rejection, determinism by seed, quantile ordering, analytic correctness on a hand-checkable linear case, failure-count behavior, all three distribution types, multi-target runs, and performance sanity

Test plan

• python -m pytest tests/test_uncertainty.py -q -- 35 passed
• python -m pytest -q -- 705 passed (670 baseline + 35 new)
• python -m gpu_stack.cli audit --fail-on-issues -- PASS
• End-to-end driver: cost_per_token with 3 synthetic uncertain inputs (price peak, price offpeak, cluster availability), n_samples=200, seed=42 -- mean 3.27e-6, std 1.46e-7, p5-p95 range 3.06e-6 to 3.50e-6, 0 failures
• Code review findings fixed: removed dead _is_finite_float helper, fixed dead if not _has_numpy: pass branch and redundant import, fixed population vs sample std (now Bessel-corrected), added duplicate variable name validation, corrected return type annotation on _try_lambdify_path

https://claude.ai/code/session_01Eu2JVnPFgMQftwYTP3cGQZ
EOF
)

---

Generated by Claude Code