fix: sync upstream + revive AdaBoost & ResidualChain (Object=M pattern, closes #3)

[jagan-nuvai]

Summary

Two concerns bundled (originally just the upstream sync, then expanded with the structural fix):

1. Sync with upstream rust-ml/linfa master (7 commits since 0.8.1)
2. Revive AdaBoost and ResidualChain by reworking their trait bounds with the Object = M pattern, eliminating the trait-solver recursion under ndarray 0.17 (closes linfa-ensemble: AdaBoost trait bounds need structural rework (orphaned in #2) #3)

If preferred, I can split into two separate PRs (sync-only first, then rework). Bundled here for atomic review since the rework un-orphans what the sync orphaned.

Part 1 — Upstream sync (commit 47c7ed1)

7 incoming commits from rust-ml/linfa master:

Commit	Description
1abc88f	feat: add symmetric mean absolute percentage error (sMAPE) (rust-ml#437)
12c6c73	fix: realign PreprocessingError variants with error strings (rust-ml#434)
17f8696	Fix label ordering in binary logistic regression (rust-ml#432)
c7c2af5	Add generic ResidualChain composing method (rust-ml#430)
b1f9ddb	Relax required test score
2197362	Update to Zola 0.22
1de164b	feat(linfa-tsne): update to bhtsne 0.5.4 (rust-ml#429)

4 conflicts resolved (3 Cargo.toml regions + 1 lib.rs autotraits test that took upstream's L2Dist API change).

Part 2 — AdaBoost + ResidualChain rework (commit 2250107)

Both wrappers used where P: Fit<...> + Clone plus a direct P::Object: SomeTrait bound, which under ndarray 0.17 cascaded into P: ParamGuard not satisfied → infinite <<P::Checked>::Checked>::Checked: ParamGuard regress.

The proper fix mirrors the working EnsembleLearnerValidParams template (linfa-ensemble/src/algorithm.rs:137): introduce a fresh generic M for the inner model type and bind via Fit<..., Object = M>. This decouples the trait solver's job into independent linear obligations:

• "P implements Fit with Object = M" (verified by where-clause assertion)
• "M implements " (independent constraint)

Instead of the chained projection that forces blanket-impl resolution.

AdaBoost (algorithms/linfa-ensemble/src/adaboost.rs):

impl<D, T, P, M, R> Fit<Array2<D>, T, Error> for AdaBoostValidParams<P, R>
where
    D: Clone + ndarray::ScalarOperand,
    T: FromTargetArrayOwned + AsTargets,
    T::Elem: Copy + Eq + Hash + std::fmt::Debug + Into<usize>,
    T::Owned: AsTargets<Elem = <T as AsTargets>::Elem>,
    P: Fit<Array2<D>, T::Owned, Error, Object = M>,
    M: PredictInplace<Array2<D>, T::Owned>,
    R: Rng + Clone,
    usize: Into<T::Elem>,

ResidualChain (src/composing/residual_chain.rs):

impl<F1, F2, M1, M2, F: Float, D: Data<Elem = F> + RawDataClone, T, E1, E2>
    Fit<Arr2<D>, T, ResidualChainError<E1, E2>> for ResidualChain<F1, F2, F>
where
    Arr2<D>: Records,
    F1: Fit<Arr2<D>, T, E1, Object = M1>,
    F2: Fit<Arr2<D>, Array1<F>, E2, Object = M2>,
    for<'a> M1: Predict<&'a Arr2<D>, Array1<F>>,
    T: AsTargets<Elem = F, Ix = Ix1>,
    E1: std::error::Error + From<crate::error::Error>,
    E2: std::error::Error + From<crate::error::Error>,

No public API change: users still write AdaBoostParams::new(DecisionTree::params()).fit(&train) and params1.chain(params2).fit(&train) exactly as before. The M generic is inferred from the call site.

Sources/tests/example restored

• algorithms/linfa-ensemble/src/lib.rs — mod adaboost; re-added, 6 AdaBoost test fns un-gated from #[cfg(any())], doc comment AdaBoost section restored.
• algorithms/linfa-ensemble/examples/adaboost_iris.rs — restored from upstream master (134 lines, no migration needed).
• src/composing/mod.rs — pub mod residual_chain; re-added, doc comment "four composition models" restored.

Verification on x86_64-linux (knuckles-dev VM)

Check	Result
cargo check --workspace --all-targets	exit 0, 0 warnings
cargo test -p linfa-ensemble --lib	16 passed (6 AdaBoost tests now active alongside 10 existing)
cargo test --workspace	running, will update once complete

Notes for reviewer

• Issue linfa-ensemble: AdaBoost trait bounds need structural rework (orphaned in #2) #3 closes: both AdaBoost and ResidualChain are revived with the rework. The Object = M pattern is what actually works — the failed alternatives explored are documented in commit 2250107's message body for posterity.
• Pattern worth documenting upstream (rust-ml/linfa): wrappers that take generic Fit-able params should bind Fit<..., Object = M> rather than P::Object: .... Saves the next contributor from the same trap. Optional follow-up: contribute a CONTRIBUTING.md note or wrap-pattern docs.
• Single bundled PR is offered for atomicity (the rework's diff makes no sense without the sync's orphan being present). Splitting is doable but adds chained-PR complexity.

Closes #3

[jagan-nuvai]

Final test verification — full workspace

cargo test --workspace on x86_64-linux (knuckles-dev VM): 509 passed, 0 failed (511s wall time).

Net delta vs. pre-rework state:

• +6 AdaBoost tests (un-gated from #[cfg(any())], now active)
• +~32 ResidualChain tests (pub mod residual_chain; re-added, all passing)
• Total: +38 tests restored to the suite

The Object = M rework is fully validated:

• Compile-time: cargo check --workspace --all-targets zero warnings
• Run-time: all six AdaBoost iris/synthetic tests + all ResidualChain doctests/integration tests + every other test in the workspace passes

Ready for review/merge.