Seed the NumPy RNG used for initial weight noise in reweight()

[MaxGhenis]

Summary

Finding #3 (HIGH). Calibration.__init__ called torch.manual_seed(self.seed) but never np.random.seed(self.seed), yet reweight() drew its initial weight noise from the unseeded global np.random stream. Two runs with the same seed therefore produced different initial log-weights and diverged thereafter, breaking the documented reproducibility guarantee and destabilising CI, hyperparameter tuning, and holdout robustness.

reweight() now accepts an explicit seed parameter. When provided it seeds torch (including CUDA) and draws initial noise from a local numpy.random.default_rng(seed), which keeps the caller's global numpy RNG state untouched. Calibration.calibrate() threads its self.seed through.

Test plan

• Add tests/test_seed_determinism.py: identical seeds → identical weights, different seeds → different weights, calibration does not mutate the caller's global numpy RNG stream.
• All existing tests pass (uv run pytest tests -x -q -> 18 passed).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
microcalibrate	Ready	Preview, Comment	Apr 17, 2026 4:23pm

[MaxGhenis]

LGTM (cannot self-approve; posting as comment).

• Local np.random.default_rng(seed) — does not mutate the caller's global numpy state. The dedicated test test_calibration_does_not_mutate_global_numpy_state verifies this.
• Explicit seed parameter on reweight(), threaded from Calibration.calibrate() via self.seed. Clean API.
• torch.manual_seed(seed) + torch.cuda.manual_seed_all(seed) guarded by if seed is not None — preserves the previous non-deterministic behaviour when no seed is given.
• Three tests (same-seed determinism, different-seed divergence, global-state preservation) cover the reproducibility contract end-to-end with noise_level > 0 so the numpy RNG is actually exercised.

Note: this branch doesn't include the np.maximum(..., 1e-12) guard from #91; once #91 merges, #93 will need a trivial rebase. That's the case for all 7 branches as they're independent off origin/main.