feat: transformers as UDAFs v0 — just run the pipeline

[claude-agent-ahrzb]

Loop 5, scoped per the design dialogue: no desugaring (the optimization pass comes later), no sklearn.* namespace (needs the sklearn index) — transformers are opaque objects and fit literally runs them. Spec: docs/superpowers/specs/2026-07-29-transformers-run-the-pipeline-design.md.

embed = Pipeline([("sc", StandardScaler()), ("pca", PCA(2))])   # found in caller scope
p = SQLProjection(
    "SELECT embed(struct_pack(a := age, f := fare)) OVER (PARTITION BY country) AS e,"
    " age - avg(age) OVER () AS c, name FROM __THIS__"
).fit(train)
out = p.transform(table)      # NEW: batch apply

• Resolution: explicit transformers= registry, then caller scope (FROM df idiom), guarded by duckdb_functions() so a local variable can never shadow real SQL. Captured objects surface on p.transformers.
• Plan: a second step kind — FitStep(kind="fit", transformer, features, keys) reading the level table; the executor grows one branch (groupby keys → clone() → fit). Everything else (names, reads, DAG order) unchanged.
• Apply: transform(table) runs serving_sql (helper feature/key columns ride along), applies fitted clones per group, splices each output column in place of its helper block. Unseen group → NULL, the exact-join policy.
• Gate: transformer columns vs an independent clone-per-group sklearn reference; SQL columns still bit-exact vs DuckDB. Named refusals for ORDER BY/frames/FILTER on transformer windows, non-final levels, nested calls, fn(*) (the parser erases the star — refused loudly rather than silently fitting zero features), unnamed struct fields, and namespaced names.

Root pytest 507 passed + 1 xfailed; confit untouched.

🤖 Generated with Claude Code