Task: Implement Hotfix: stochastic process statistical tests fail on macOS CI

Widen the statistical tolerances in the two failing tests so they hold for any platform's std::normal_distribution sequence while still bounding genuine model errors:

• black_karasinski_process_tests.cpp (kappa <= 0 walk test): sample variance of 200k paths must match sigma^2 * dt; the 3e-3 epsilon allows only ~0.95 sigma of the estimator's standard error.
• heath_jarrow_morton_process_tests.cpp (drift grid-integral test): per-tick drift means must match the hand-computed vector; the 2e-3 margin is ~0.6 sigma for the largest-vol rate.

• Both tests pass on Linux with the current libstdc++ sequence.
• Tolerances are >= ~3 sigma of the estimator's standard error, so the libc++ (macOS) draw sequence cannot land outside them.

Root cause: std::normal_distribution is implementation-defined; libc++ and libstdc++ map the same mt19937 output to different normal draws. The tests seed each path (seeds 42+i) and were tuned against the Linux sequence, with tolerances tighter than the tests' own standard error:

• BK variance test (lines 240-267): N = 200000 paths; standard error of the sample variance is sigma^2 * sqrt(2/N) = 1.42e-4, but the 3e-3 epsilon admits only 1.35e-4 (0.95 sigma). macOS draws gave 0.044853 vs the expected 0.045.
• HJM drift test (lines 163-194): N = 100000 paths; per-tick drift mean for rate 3 has standard error ~3.2e-3, but the margin is 2e-3 (0.6 sigma). macOS draws gave mean[1] = 0.04733 vs 0.045.

Fix: widen the bounds to multi-sigma headroom. BK epsilon to 1.5e-2 (the sibling Gaussian-OU test in the same file already uses 1.5e-2 for the same estimator at the same path count); HJM margin to 1e-2. This makes the failure probability per platform negligible while keeping the checks ~4 sigma tight relative to genuine deviations.

Root cause: the two Monte Carlo tests assert tolerances tighter than their own standard error, and std::normal_distribution maps the same mt19937 seeds to different draws per standard library, so the macOS (libc++) sequence lands outside the bounds the tests were tuned to on Linux (libstdc++).

Fix (committed): widen the bounds to multi-sigma headroom, matching the sibling conventions in the same files:

• BK kappa <= 0 walk test: variance epsilon 3e-3 → 1.5e-2 (~4.7 sigma at 200000 paths; the Gaussian-OU moments test already uses 1.5e-2 for the same estimator and path count).
• HJM drift test: mean margins 2e-3 → 1e-2 (>= 3.2 sigma at the noisiest rate, sigma = 0.5; the previous margin was ~0.6 sigma). The front-rate zero-mean check gets the same margin.

Verification: `compass build ores.analytics.quant.tests` green; running `[black_karasinski_process],[heath_jarrow_morton_process]` passes all 34 test cases (2084209 assertions) on Linux. No source changes — test-only hotfix; a genuine model error of ~3 sigma or more is still rejected.

Manual QA scenarios (scaffolded via compass add test_scenario, run through the QA Validation Runner panel) that verify this task. Link new ones here as they're created; the scenario doc itself links back via its "Verifies task" field.

Fix verified locally (build green, both suites pass: 34 test cases, 2084209 assertions) and shipped as PR #1976. Review round: one minor observation on per-rate HJM margins, declined with rationale (error and standard error both scale with the rate's volatility, so detection is leg-independent in sigma units); reviewer approved, claude-review and check gates green.