linfa-ensemble: AdaBoost trait bounds need structural rework (orphaned in #2)

[jagan-nuvai]

Summary

The impl<D, T, P, R> Fit<Array2<D>, T, Error> for AdaBoostValidParams<P, R> block in algorithms/linfa-ensemble/src/adaboost.rs cannot satisfy Rust's trait solver under our ndarray 0.17 / rand 0.9 ecosystem. AdaBoost source files have been temporarily orphaned (kept on disk, not declared via mod in lib.rs) on the feat/ndarray-0.17-upgrade branch (PR #2) so the dependency upgrade can ship.

Repro on master (post-rebase onto ndarray 0.17)

cargo check -p linfa-ensemble --lib produces ~26 cascading errors of the form:

• the trait bound \P: ParamGuard` is not satisfied`
• After adding P: ParamGuard: <P as ParamGuard>::Checked: ParamGuard is not satisfied
• After adding that: <<P as ParamGuard>::Checked as ParamGuard>::Checked: ParamGuard is not satisfied
• And so on — infinite regress.

Root cause

Linfa's Fit blanket impl in src/param_guard.rs:55-66:

```rust
impl<R: Records, T, E, P: ParamGuard> Fit<R, T, E> for P
where
P::Checked: Fit<R, T, E>,
E: Error + Fromcrate::error::Error + From<P::Error>,
```

AdaBoost writes where P: Fit<Array2<D>, T, Error> — for generic P, the only way to satisfy this is via the blanket. The blanket then needs P::Checked: Fit, which (for unknown P::Checked) again only resolves via blanket → needs P::Checked: ParamGuard → recursive.

This may have worked under older nightly / older ndarray because of slight differences in trait selection / coherence checks. Under stable + ndarray 0.17 it doesn't.

What's been done in PR #2

• mod adaboost; and mod adaboost_hyperparams; removed from linfa-ensemble/src/lib.rs.
• 6 AdaBoost test functions gated with #[cfg(any())] (always-false; preserves source).
• Doc comment AdaBoost section removed.
• examples/adaboost_iris.rs deleted.
• Source files src/adaboost.rs, src/adaboost_hyperparams.rs left on disk (orphaned). Cargo doesn't see them since they're not declared via mod.

Suggested fix paths

1. Structural rework: Replace P: Fit + Clone with a model factory:

   ```rust
   pub struct AdaBoostValidParams<F, R> {
   model_factory: F, // Fn(&BootstrapDataset) -> Result<Box<dyn Predict<...>>, Error>
   n_estimators: usize,
   ...
   }
   ```

   Sidesteps trait-system entanglement entirely.

2. Use P::Checked: Fit directly + manual check call: change body to call self.model_params.clone().check()?.fit(...) and bound where P: ParamGuard, P::Checked: Fit<...>, Error: From<P::Error>. Does NOT work today (compiler still demands <P::Checked>::Checked: ParamGuard); would need either trait-system improvement upstream or a workaround.

3. Erase the trait genericism: take &dyn Trait or specific concrete model types only. Less flexible but compile-stable.

Acceptance criteria

• cargo check -p linfa-ensemble --lib warning-free
• cargo test -p linfa-ensemble --lib — 6 AdaBoost tests un-gated and passing
• examples/adaboost_iris.rs restored
• mod adaboost; / pub use adaboost::*; restored in linfa-ensemble/src/lib.rs
• Original PR Add AdaBoost classifier to linfa-ensemble rust-ml/linfa#427's intent preserved (AdaBoost still wraps any Fit-able weak learner)

References

• Original PR: Add AdaBoost classifier to linfa-ensemble rust-ml/linfa#427 (AdaBoost classifier)
• Blocking PR: fix: complete rand 0.9 + ndarray 0.17 migration (closes #1) #2 (ndarray 0.17 / rand 0.9 migration)
• Commit that orphaned: 09b2f39 ("Merge origin/master into feat/ndarray-0.17-upgrade")
• Related upstream issue if you find one: rand-0.9 / ndarray-0.17 + Linfa's blanket Fit interaction

[jagan-nuvai]

Update — second instance of this pattern hit in PR #4

Upstream rust-ml/linfa added ResidualChain (PR rust-ml#430, commit c7c2af5) — a generic wrapper that fits models on the residuals of previous models. Its impl Fit for ResidualChain<F1, F2, F> uses the same where F1: Fit + ParamGuard, F2: Fit + ParamGuard shape that triggered the AdaBoost cascade.

When syncing this upstream change in PR #4, ResidualChain produced an identical 15-error cascade:

• F1: ParamGuard not satisfied
• F2: ParamGuard not satisfied
• <F1 as ParamGuard>::Checked: ParamGuard not satisfied (infinite regress)

PR #4 orphaned ResidualChain using the same pattern as AdaBoost (removed pub mod from src/composing/mod.rs, source preserved on disk).

This means the structural rework owed under this issue covers two call sites now:

1. algorithms/linfa-ensemble/src/adaboost.rs (AdaBoost)
2. src/composing/residual_chain.rs (ResidualChain)

Suggesting we widen the title of this issue to "linfa: generic-wrapper-around-Fit pattern triggers trait-solver recursion" since this is now visibly a systemic API style issue, not a one-off in AdaBoost.

The fix paths from the original issue body still apply — model-factory / take-a-Fit-Object directly / dyn Trait erasure all sidestep the ParamGuard blanket entanglement. Whichever rework path is chosen, both AdaBoost and ResidualChain should be migrated together since they're isomorphic.