fix(orm): grouped where(lambda) prefixes subgroup columns with the table (GH #205)

[tmgbedu]

Fixes #205.

Problem

A grouped where(lambda q: ...) rendered invalid SQL — the subgroup's columns lost their table prefix:

Message.where("thread_id", "t").where(
    lambda q: q.where("role", "user").where("type", "text").or_where("id", ">=", 42)
).to_sql()
# ... AND (."role" = 'user' AND ."type" = 'text' OR ."id" >= '42')

Postgres rejects ."role" with syntax error at or near ".".

Root cause

QueryBuilder.new() returned self.connection.query() — a builder with _table = "" and no model — and where() builds the subgroup as column(self.new()). The grammar then rendered the subgroup columns with an empty table prefix.

Fix

new() now carries the current table into the nested builder:

def new(self):
    return self.connection.query().table(self._table)

• The where(lambda ...) subgroup inherits the parent table, so columns render as "messages"."role" etc.
• Callers that need a different table (e.g. BelongsToMany's exists/join subqueries) already chain .table(...) right after new(), so their behaviour is unchanged.
• The model / global scopes are intentionally not copied, so those executed subqueries stay clean (no accidental scope leakage).

Tests

New regression tests in tests/masoniteorm/sqlite/builder/test_sqlite_query_builder.py covering the exact grouped AND/OR pattern from the issue:

• to_sql() → ... AND ("users"."active" = '1' AND "users"."age" > '20' OR "users"."id" >= '42')
• to_qmark() → correct ? placeholders and binding order ['Joe', 1, 20, 42]

Full ORM suite (excl. live-Postgres): 723 passed, 7 skipped. The generated SQL is now valid Postgres, so the CI Postgres job exercises it end-to-end.

Real-world validation (follow-up)

The where_raw workaround this unblocks lives in example/agents/app/repositories/conversation.py on the ai-package-fix branch (not on main), so its revert isn't part of this PR. The added regression test replicates that exact ("role" = ? AND "type" = ? OR "id" >= ?) shape; once this lands and ai-package-fix syncs, that line reverts to where(lambda q: q.where("role","user").where("type","text").or_where("id",">=",boundary)).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[tmgbedu]

✅ Code Review: APPROVE (posted as comment — GitHub blocks self-approval on this account)

Reviewed the diff, traced every new() caller, and exercised edge cases beyond the two new tests.

Correctness

new() now returns self.connection.query().table(self._table). The grouped-lambda path (column(self.new()) in where()) makes the nested builder inherit the parent table, so subgroup columns render as "users"."col" instead of the invalid table-less ."col" Postgres rejects.

Binding order

to_qmark() yields bindings in source order ['Joe', 1, 20, 42] — subgroup wheres are appended in-order into the parent _bindings; grouping doesn't reorder. Covered by the new qmark test.

Deliberate NOT-copying model/global scopes — confirmed safe

• Pre-fix new() already returned a bare builder with no model, no scopes; this change only adds .table(), so scope/model behavior is unchanged — nothing newly omitted.
• Correct for a where(lambda) subgroup: it only renders inline _wheres into the parent (never executed standalone), so re-applying global scopes there would be wrong. Not copying keeps scoped subqueries clean.
• Both BelongsToMany callers (query_where_exists, query_has) chain .table(table) right after new(), overriding the carried table — behavior unchanged.

Edge cases I verified manually

• Explicitly-qualified column in a subgroup is not double-prefixed: where('roles.name',…) → "roles"."name".
• Nested subgroups inherit the table at every level: ("users"."a" = '1' AND ("users"."b" = '2' OR "users"."c" = '3')).

Tests & CI

• test_sqlite_query_builder.py → 25 passed (incl. 2 new regression tests). Full tests/masoniteorm → 726 passed, 7 skipped; the 17 failures are all local live-Postgres (role "app" does not exist) — unrelated. CI green: Pytest, Ruff, Pyright all pass; mergeState CLEAN.

Regression coverage matches the issue precisely. Recommend merge — holding for QA sign-off (task #1400) per PM instruction.