Add LoadSharingModel: load-sharing dynamic node (#38)

[derrynknife]

Summary

Adds LoadSharingModel (#38): a dynamic RBD node for load-sharing / dependent failure, a sibling to StandbyModel. n coupled units share a total load; when one fails the survivors carry more (L / s) and — through their AFT phi(load) response — age faster. The group works while ≥ k of n survive. Because the units are coupled, the group is a single node (it can't be n independent nodes).

Engine — a cumulative-exposure event loop (per Monte-Carlo replicate): draw each unit's baseline exposure-to-failure threshold, advance every survivor's baseline clock at rate phi(L / survivors), drop the first to reach its threshold, repeat until fewer than k survive.

This is the "self-loading" sibling of the condition-based layer: same conditional-survival engine, but the load comes from the group's own survivors rather than from sensors.

• New repyability/rbd/load_sharing_node.py: sf/ff/mean/random/cs. Identical Exponential-baseline units get an exact hypoexponential closed form (stage rates s·λ·phi(L/s)); otherwise the survival is a Kaplan-Meier fit to simulated lifetimes (is_simulated reports which).
• Wired into NonRepairableRBD exactly like StandbyModel (simulation node, time-varying, non-stateable, node_mttf), RBD (de)serialisation, and exported from the package.

Type of change

• Bug fix
• New feature
• Documentation
• Refactor / tooling
• Breaking change

Checklist

• Tests added/updated (asserting against a reference value where possible)
• black, isort, flake8, and mypy pass
• coverage run -m pytest passes and stays above the coverage gate
• CHANGELOG.md updated under [Unreleased]
• Any Monte-Carlo tests are seeded (deterministic)

Notes for reviewers

• Exactness anchor: with no load effect (phi ≡ 1) the stage rates become s·λ and the closed form is exactly the (N-k+1)-th order statistic of N i.i.d. Exponentials — the ordinary k-out-of-n parallel result. A test checks this against the binomial-sum formula, and closed-form vs Monte-Carlo agreement is tested for the load-dependent case.
• AFT-specific by design: the exposure clock relies on the AFT time-scale property, so units are fitted surpyval AFT models fitted with load as the covariate. This is the same phi/exposure machinery kept internal here — it does not touch the public RegressionNode/NodeState API.
• Distribution fitting stays in surpyval; RePyability consumes the fitted models.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NUGaeQXKxjjiQ1ULT1ApXS

---

Generated by Claude Code