Board-only change (backlog/**). One new draft, no code.

DRAFT-14 — transformer(x) OVER (PARTITION BY g) fits one transformer per group

AmirHossein's proposal, routed via Wren out of TASK-3 brainstorming. Wren's measurements are what make it scopeable, so they're recorded verbatim in the ticket.

The reframe that matters: today a transformer ref is fitted once globally — measured, with country a={10,20}, b={30,50} the state holds avg_age=[27.5], not per-country 15/40. That is the correct semantic (it matches sklearn, where a Pipeline step fits once on all training data), so this is a new feature, not a bug.

The split, which is the spine of the ticket:

• (a) SQL-expressible (StandardScaler, MinMax, MaxAbs, mean-SimpleImputer) — already works end-to-end today. Pure desugar, zero engine work. Measured: the PARTITION BY form yields one state row per group, batch == infer, and unseen groups return NULL for free via the existing unseen-partition semantic.
• (b) Opaque sklearn objects — genuinely new. N fitted clones keyed by group, per-group state through the artifact and lookup path.

Four open questions, why it's a draft not a task:

1. Unseen-group policy for (b) — NULL, global fallback, or error? doc-2 calls unknown-category handling "a designed-in requirement, not a flag." Wren's point: (a) inherits NULL by construction, so any other choice for (b) creates a semantic divergence between two halves of one feature.
2. Syntax — OVER (g) shorthand vs standard OVER (PARTITION BY g). Wren leans standard; (a) then desugars into already-valid SQL, whereas a shorthand buys new parsing for no new capability.
3. Ship (a) alone first? Nearly free, covers the common case.
4. Artifact size for (b) — N clones grows the artifact with group cardinality.

Explicitly flagged as distinct from DRAFT-11: that one is about how arguments bind; this is about what data a transformer is fitted on.

🤖 Generated with Claude Code