Simplification: medium-risk numerical dedup batch — AJ increment helper, risk-set convention helper, bootstrap_cb merge, CB transform sharing, Cox TVC wrappers, CopulaModel serialisation

[derrynknife]

Remaining medium-risk items from the simplification review, grouped:

1. Aalen-Johansen increment implemented three times: competing_risks/nonparametric/competing_risks.py:245-248, competing_risks/regression/competing_risks_proportional_hazard.py:110-120 (same Competing risks: CIF uses S(t) instead of S(t-); queries before first event wrap to last value; fit_from_df shadows df() #253/Competing risks: default Nelson-Aalen method produces CIFs that exceed 1 #278 comments), gray_test.py:190-205 (loop form). One aalen_johansen_iif(S_km, hazard_increments) helper; Competing risks: default Nelson-Aalen method produces CIFs that exceed 1 #278 was exactly one of these copies going wrong.
2. Cox at-risk rule (tl < tau) & (x >= tau) hand-rolled in three Python loops (cox_ph.baseline:289-295 — O(K·N), _prepare_exact_tie_data:519-527, diagnostics._risk_set_means:102-105) while vectorised equivalents (at_risk_beta_Z + not_yet_entered) exist in the same file. Minimum: vectorise baseline(); add one documented risk-set-convention helper so the entry-strict/exit-inclusive rule lives in one place. Pinned by the R/lifelines comparison tests.
3. degradation/_bounds.py bootstrap_cb vs bootstrap_cb_accelerated (~40 lines): same resample→refit→quantile loop; merge with Z=None.
4. CB transform sharing: parametric.py:1394-1465 and regression/_bounds.py:48-116 implement the same logit-sf / log-rate transforms with different jacobian engines (autograd vs FD — keep both engines, share only the transform functions and align the two-sided output layout).
5. Cox TVC entry-point quartet (cox_ph.py:955-1132) structurally mirrors TVCFitMixin (tvc_fit.py:42-148); a _post_tvc_fit hook collapses at least the two *_from_df wrappers (Cox's docstrings are worth keeping, which limits the net win).
6. CopulaModel.to_dict (multivariate/parametric/copula/copula_model.py:89-99) is the only to_dict without stamp_schema, has no from_dict, and writes a "parameterization" tag surpyval.from_dict doesn't recognise — it serialises to something nothing can read. Either register + implement from_dict, or document one-way export.

Each item should land with before/after numerical-equality checks on the relevant suites. (Low-risk simplification items from the same review are being applied directly.)

[derrynknife]

Items 1–3 and 6 (Aalen-Johansen increment helper, cox_at_risk_mask + vectorised baseline(), degradation bootstrap merge, CopulaModel round-trip serialisation) are implemented in PR #301.

Items 4 (CB transform sharing between parametric.py and regression/_bounds.py) and 5 (Cox TVC *_from_df wrapper collapse) are deferred — they have the least net win for the churn involved. Keeping this issue open for those two.

---

Generated by Claude Code