[derrynknife]

Fixes the failures in the first full run on the release PR (#984). None of them showed up locally on arm64 macOS.

• Regression qf (test_571_qf_inverts_ff_for_every_family[LogNormalPH], py3.12/3.13).
  • Newton's finish rejected every step when the bracketed answer sat at an end of its bracket and the cumulative hazard's sign there flipped between batches. numpy 2.5.4's vectorised exp/log on Linux round the same time differently in a batch of another size.
  • So qf(0.9) kept the solver's 1.1e-12 error.
  • The fix widens the bracket by 1e-10·|t|; a step must still reduce |gap|.
  • Reproduced in Linux containers (arm64 and amd64) and fixed there; the error is now 4e-16.
• test_828_qf_is_exact_near_the_start_of_the_support[NormalAH].
  • The check bounded ff(qf(s)) near an additive model's support start by 1e-8 of s. That is about two ulps of the terms H0 + x·β'Z, which cancel there, and x86-64 lands at 1.1e-8.
  • The bound is now 16 ulps of |H0(q)|, the limit any qf or ff has there. The largest seen is about 5 ulps, on mac, Linux arm64 and Linux amd64.
• test_standard_errors_match_the_numerical_ones[NormalAH].
  • The exact SEs agree across platforms; it is the numerical reference that drifts: 1.05e-4 from the exact SEs on x86-64, 6.2e-5 on arm64.
  • NormalAH joins OLD_ROUNDING (2e-4), which also checks at ten times the step, where they agree to 6.7e-7.
• Doctests (Parametric.plot, both plot_partial_effects).
  • CI runs doctests sequentially in one process, so an example drawing on plt.gca() drew on an earlier example's axes.
  • The root conftest now closes all figures before each doctest. Reproduced locally with a sequential run: 3 failed before, 805 passed after.
• Refit calibration registry. WeibullJointPH was neither planned nor excluded. It is now excluded, because its own study is calibration/test_coverage_joint.py.

Local: regression + fast conformance 8462 passed; sequential doctests 805 passed; lint clean (pinned pre-commit isort and isort 9.0.2).

🤖 Generated with Claude Code

https://claude.ai/code/session_019Km2eWBiPJ5JNkme3vRg4B