nirs4all-methods — gaps for the in-browser studio-lite (WASM)

nirs4all-lite/studio-lite is a full-WASM browser app that drives libn4m through the emscripten convenience surface (bindings/js/src/wasm_entry.c, the _n4m_wasm_* helpers) plus a curated catalog (nirs4all-lite/studio-lite/src/catalog/nodes.ts, CI-gated against cpp/abi/expected_symbols_*.txt). The items below are operators/capabilities that exist in libn4m but are not yet usable from the browser, so studio-lite either omits them or falls back. None require changing the public ABI snapshot — the wasm helpers are emscripten-only _n4m_wasm_* exports appended in bindings/js/CMakeLists.txt.

This is a forward-development backlog, not a bug list.

0. Held commits to reconcile with the in-flight AOM work

Two additive bindings/js commits are held on release-readiness-fixes (NOT pushed) while the AOM core is being developed, to avoid colliding:

  • 8d2371e — MIR-PLS / MB-PLS / missing-aware NIPALS added to the generic model dispatcher.

  • 4fc0a7a — split-operator wasm dispatcher (Kennard-Stone / SPXY / KMeans / KBins-stratified). The deployed demo runs the staged wasm built from them. When AOM settles: reconcile + push (or rebuild + re-stage the wasm against the finished AOM).

1. Non-coefficient model predict surface (blocks ~7 models in the editor)

The generic dispatcher n4m_wasm_model_fit only supports models that expose an input-space coefficient triple (coefficients + x_mean + y_mean [+ real intercept]) so the browser predicts via the centred/affine form (n4m_wasm_model_predict_from_coeffs). These exported models don’t fit that contract, so they’re absent from the editor:

model

fit symbol

why it doesn’t fit

needs

PLS-LDA / PLS-QDA / PLS-logistic

n4m_pls_lda_fit / _qda_fit / _logistic_fit

classification heads expose decision scores / score-space, not input-space coeffs

a wasm decision_function / predict_proba-style predict

NPLS / ONPLS

n4m_n_pls_fit / n4m_on_pls_fit

multi-way (tensor) / loadings only

tensor input contract + predict

GPR-PLS

n4m_gpr_pls_fit

returns posterior/variance

predict returning mean (+ optional variance)

LW-PLS

n4m_lw_pls_fit

locally weighted — per-query, no global coeffs

“fit → portable state + predict(X_new)” surface

Kernel-PLS

n4m_kernel_pls_fit

dual coefficients

kernel-predict surface (store support + kernel params)

DI-PLS

n4m_di_pls_fit

needs a target domain

target-domain input

Proposed fix: a generic n4m_wasm_model_predict(model_token, fitted_state_blob, X_new) → Y so the browser no longer needs the coeff-triple shortcut; each model serializes a portable fitted state the predict path consumes.

2. Preprocessing dispatcher — shape-changing + stateful gaps

The generic n4m_wasm_pp_* dispatcher only handles shape-preserving (n×p → n×p) transforms, and only stateless ops + MSC (which got internal *_get_state/*_set_state).

  • Shape-changing ops (crop, resample, edge-trimming derivatives, PCA/SVD feature projection, wavelet features, interval selection) need an output-shape query (n4m_wasm_pp_out_cols(...)) + a non-shape-preserving transform path. The catalog excludes them today.

  • Stateful ops beyond MSC (EMSC, EPO, OSC, direct standardization, baseline-with-fitted-reference, COW/DTW alignment) need MSC-style internal state accessors (*_state_n / *_get_state / *_set_state) so the fitted state round-trips for predict-later (.n4a). Without them they can’t be exposed safely.

3. AOM / POP — richer inputs

  • The operator-bank is exposed, but WeightedPLS (needs a length-n weights vector) and multi-group GroupSparsePLS (needs feature group-ids) need an auxiliary array param the scalar NodeDef can’t carry — a wasm fit variant taking the extra array.

  • AOM branch/selection modes beyond the duplication-style screen (by_source) need multi-source X (see also dag-ml’s multi-source gap).

4. Moments engine (infra — not an editor operator)

n4m_moments_* (μ, XᵀY, XᵀX) is infrastructure used inside AOM/PLS sweeps; it maps to no user-facing editor operator. Not needed for the editor. A n4m_wasm_moments_* surface would only matter if the in-browser “moment engine” tiers become a goal (cf. the gpu-moment-engine plan).

Source: the studio-lite build (2026-06). The browser/WASM binding is the layer to grow; the C++ core already implements the numerics.