boost: reconcile nns_boost with R 13.2 NNS.boost; version 2.0.0

[OVVO-Financial]

Summary

Port of the reconciled R NNS.boost (OVVO-Financial/NNS PR #58, commit bb878ad), regenerated parity cache, restored automation fixes, and the 2.0.0 major bump.

nns_boost reconciliation

• threshold is a probability supplied to lpm_var over the learner-trial objective distribution (defaults 0.80 max / 0.20 min; extreme selects 1/0) — never a literal score cutoff. Survivors and every epoch re-test use the distinct LPM.VaR objective cutoff.
• Learner trials fit genuine multivariate nns_reg (dim_red_method=None) and draw a fresh validation holdout per cross-sectional trial; time-series trials keep the chronological terminal block.
• Epochs always run (exhaustive enumeration no longer disables them), re-testing only surviving combinations with R's survivor-permutation scheduling and fresh holdouts (chronological expanding windows for ts).
• Final estimate replicates the original keeper predictors by relative epoch frequencies into one multivariate nns_stack(method=1, stack=False, folds=5, balance=False) call; the synthetic X* path is removed. New public folds argument (default 5).
• Result gains class.levels (None for regression), matching R.

Every draw replicates R's Mersenne-Twister stream via RRNG — all seeded paths are bit-for-bit identical to R, including balance=True, which closes the formerly xfail-characterized balanced-boost sampling gap (the iris classification vignette now matches installed R exactly; its gap test pins the closure).

Cache

The three NNS.boost.* shards are regenerated from installed R NNS 13.2 at bb878ad (the automation run lost them — see below); sync/nns_source.json records that commit. All 37 boost parity tests pass bit-for-bit.

Restored workflow fixes (needs workflow-change-approved)

The #136 workflow improvements (sharded add-paths, shard-aware validation, manifest provenance stamping, per-function cache-diff table in the PR body, CI matrix dedupe) were silently reverted by the workflow-baseline protection, which caused the post-#58 cache regeneration to commit nothing and delete its own branch. They are restored here — this PR needs the workflow-change-approved label before merge or the baseline will strip them again.

Version 2.0.0

The threshold argument's meaning changed on an unchanged signature (probability, not score cutoff) and the output gains a key — semver major.

Validation

• 2098 passed cache-only across parity + invariants + tools; ruff and mypy clean; version consistency green.
• Known pre-existing, local-only failures (unchanged by this PR, skipped in CI): two Boston-housing live-R NNS.stack practical examples diverge ~0.7% from installed R 13.2.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AxBV5HmXHY5nCUvW9SwDvb

---

Generated by Claude Code