Pull Request

Short Summary

Adds a fluent experiment API — opsml.log_metric(), opsml.log_param(), and friends — backed by a process-global ActiveStack<Experiment> in Rust, with per-framework log_model shortcuts (sklearn_log_model, torch_log_model, etc.) that register a ModelCard in one call. Also refactors the trace dashboard filter state from flat fields into a composable FilterClause tree, dropping N parallel facet API calls down to 2.

Context

Fluent experiment API

Previously, logging to an experiment required a captured reference: exp.log_metric(...). This PR adds free functions that resolve the current experiment from a process-global stack pushed/popped by Experiment.__enter__/__exit__:

Before:

with start_experiment(space="training", name="run") as exp:
    exp.log_metric("accuracy", 0.94)
    exp.log_parameter("lr", 0.001)

After:

import opsml

with start_experiment(space="training", name="run"):
    opsml.log_metric("accuracy", 0.94)
    opsml.log_param("lr", 0.001)

The stack lives in crates/opsml_experiment/src/active.rs as a Mutex<Vec<Py<Experiment>>>. Single-threaded only — the comment is explicit that concurrent experiments must use the explicit form. active_experiment() exposes the current experiment for callers that need it directly.

Flavor log_model functions

Each supported framework now has a {framework}_log_model(model, *, name, space=None, ...) free function in py-opsml/src/flavors.rs. When space is omitted, it reads the active experiment's space — so opsml.sklearn.log_model(...) inside a with start_experiment(...) block just works. Returns a registered ModelCard.

ModelInterface subclass fix

PyO3 requires __new__ and __init__ to be separate for Python subclassing to work. ModelInterface previously used #[new] with a full constructor signature, which broke subclasses that pass extra kwargs. Split into __new__(*args, **kwargs) -> Self::default() + __init__(self, model=None, ...) that does the real initialization. The stubs were updated to reflect the **kwargs in __init__ and the explicit __new__ signature.

Trace filter refactor

The TraceDashboard was managing filter state as flat struct fields (service_name, service_namespace, service_version, service_instance_id, status_code, has_errors, attribute_filters). Every filter change fired up to 5 parallel getServerTraceFacets calls (one per field cleared to compute "what else matches"). This PR replaces the flat shape with a composable FilterClause tree in clause.ts:

// Before: clear individual fields per-facet call
getServerTraceFacets(fetch, { ...filters.filters, service_name: undefined })
getServerTraceFacets(fetch, { ...filters.filters, service_namespace: undefined })
// ... 3 more

// After: remove one dimension from the clause
getServerTraceFacets(fetch, { ...filters.filters, clause: removeClauseDimension(clause, "service") })
getServerTraceFacets(fetch, { ...filters.filters, clause: removeClauseDimension(clause, "status_code") })

Down to 2 calls. clause.ts (~370 lines) covers builders (serviceClause, statusCodeClause, attrClause, durationMinClause, etc.), combinators (andClause, addToClause, replaceClauseDimension, removeClauseDimension), and an ActiveFilter type for chip display. Also adds validation.ts for trace span input validation and waterfall.ts for the depth-first span sort that was previously inlined in TraceWaterfall.svelte. SvelteMap/SvelteSet replaces bare Map/Set for reactive correctness.

Dev tooling

mise.toml gains dev:e2e:* tasks for a full E2E harness against a live Scouter instance (expects a sibling ../scouter checkout or SCOUTER_DIR override). Python formatter consolidated from isort + black + ruff to ruff format + ruff check --fix. pylint dropped in favor of ty. Both Makefile and py-opsml/makefile removed.

Is this a Breaking Change?

No. All Python additions are new exports; existing Experiment method signatures are unchanged. The trace filter refactor is internal UI state — the API server routes and response contracts are unmodified. ModelInterface.__new__ now accepts *args/**kwargs, which is strictly more permissive for Python subclasses.