feat(builder): union columns + DEFAULT for mixed-shape InsertRecords

[klaidliadon]

#50 rejected any batch whose rows produced different Map column sets — ,omitzero slice / ,omitempty map records that mix nil and non-nil empty values hit this constantly, forcing callers to split into per-row InsertRecord calls. Closes #53.

records := []*Order{
    {Name: "first"},                              // cols = [name]
    {Name: "second", Tags: []string{"a", "b"}},   // cols = [name, tags]
    {Name: "third",  Note: &note},                // cols = [name, note]
}
DB.SQL.InsertRecords(records, "orders")
// → INSERT INTO orders (name,note,tags) VALUES
//   ($1,DEFAULT,DEFAULT),($2,DEFAULT,$3),($4,$5,DEFAULT)

PostgreSQL accepts DEFAULT in any VALUES position. The fix: union the cols across rows, emit sq.Expr("DEFAULT") for any slot a row skipped.

---

Design

Drafted via three Codex review rounds. Plan at tmp/insertrecords-mixed-shape/2026-06-02-plan.md (branch-only).

Key inflection points the reviews forced:

1. v1 had per-row empty-record rejection — kept feat(mapper): support ,omitzero tag option #50's behavior of erroring when any individual row mapped to zero cols. Codex blocker: that's now arbitrary. If another row contributes the union, a zero-column row is representable as (DEFAULT, DEFAULT, ...). Only the whole-batch empty union should error.
2. v1 padded by slice position — fragile. v2 pads by column name via map[string]any per row.
3. v2 hinted at SQL.InsertDefaults in the whole-batch-empty error — but that API is on the unmerged feat(builder): InsertDefaults for INSERT ... DEFAULT VALUES #52 branch. v3 hints at sq.Expr (matches feat(mapper): support ,omitzero tag option #50's existing single-row hint) so InsertRecords: union columns + emit DEFAULT for missing, instead of rejecting mixed-shape batches #53 stays independent of feat(builder): InsertDefaults for INSERT ... DEFAULT VALUES #52. Whichever lands second updates its own hint.

Reuses sqlDefault = sq.Expr("DEFAULT") from mapper.go:26. slices.Sorted(maps.Keys(colSet)) matches Map's lexical column order at mapper.go:161, so generated SQL lines up with what callers see from Map(record) directly.

Behavior table

Batch	Before (#50)	After (#53)
All rows same shape, non-empty	works	works (unchanged)
Mixed shapes, all rows non-empty	rejected (drift)	unions cols + DEFAULT fills
Some rows empty, some non-empty	rejected (drift + empty)	empty rows become all-DEFAULT
All rows empty	rejected (per-row empty)	rejected (whole-batch empty)
Empty slice of records	rejected (existing)	rejected (existing)

Test plan

• make db-reset test-all against PostgreSQL 18.3 — all 6 packages green.
• 7 unit tests: uniform shape (with args assertion), mixed shape with DEFAULT slots in the right places, ,omitzero mixed slices, ,omitempty mixed maps, empty row mixed with non-empty, all-rows-empty rejection, heterogeneous map[string]any records.
• Integration TestInsertRecordsMixedShapeRoundTrip: exec a three-row heterogeneous batch against a new dedicated mixed_shape table (nullable + defaulted columns), verify each row landed with the right mix of caller values / DB defaults / NULLs.
• go vet ./... clean.

Notes

• ,omitempty semantics preserved: rows where ,omitempty skipped a column (e.g. []string{} on a slice field) get DEFAULT in the union slot, not the empty literal. The user tagged the column for "use DB default when missing" — that contract still holds, just now inside a batch.
• Hint coordination: if #52 lands first, this PR's whole-batch-empty error could be updated to point at SQL.InsertDefaults instead of sq.Expr. Independent edit, no merge ordering required.