test: add ORM coverage for BelongsToMany and testing transaction utilities

[tmgbedu]

Rebased onto current main and trimmed to the genuinely net-new, green content.

What this PR adds

• test_belongs_to_many.py — sqlite coverage for the BelongsToMany relationship: attach/detach, attach_related/detach_related, eager loading (incl. empty), pivot access, with_timestamps, explicit pivot table, and the query-has helpers.
• testing/test_transaction.py — unit tests for the DatabaseTransaction / RefreshDatabase test-helper utilities (previously used by the test infra but never unit-tested).
• Latent-gap fixes required by the above (existing relationship code already called these but they were undefined on main):
  • QueryBuilder.table() and QueryBuilder.without_global_scopes()
  • Attribute.delete_attribute()
  • BelongsToMany pivot attach/detach refactored to build pivot rows through the query builder directly.

Dropped since this branch was opened

• The fillable-guard fix that named this branch has already landed on main (commits 82439510, 660698f8, a2d9f847, f2d82f03 are all ancestors of main) — nothing left to do there.
• Duplicate post-processor tests — main already ships tests/masoniteorm/query/processors/test_post_processors.py (merged coverage PRs) covering the same four processor classes; the near-identical copy here was removed.
• Morph relationship tests + source changes (MorphOne/MorphMany/MorphToMany) — these were written against the pre-Coverage: masoniteorm relationships (BelongsToMany, Morph*) #159 relationship internals and fail on current main. Making them pass requires completing the MorphOne/MorphToMany registry/descriptor migration that Coverage: masoniteorm relationships (BelongsToMany, Morph*) #159 applied to MorphMany only. That is a separate, reviewable change and is tracked as a follow-up rather than bundled here.

Gates

• uv run pytest --ignore=tests/masoniteorm/postgres --cov → 1820 passed, 7 skipped, total coverage 79.65% (fail_under 68%).
• ruff check + ruff format --check clean.
• No coverage regression vs main (+29 tests).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[tmgbedu]

✅ Code review verdict: APPROVE (posting as comment — GitHub blocks self-approval on own PR).

APPROVE — review only (not merging).

Reproduced gates from fastapi_startkit/:

• uv run pytest --ignore=tests/masoniteorm/postgres --cov → 1820 passed / 7 skipped, 79.65% (> 78.64% baseline). ✓
• ruff check on all touched files → clean. ✓
• Scope: exactly the 6 expected files (verified against real parent 9f01c543). ✓

Latent-bug fixes verified (methods main's BelongsToMany already calls but were undefined anywhere):

1. Attribute.delete_attribute() — called in apply_query/get_related; grep confirms no definition on main. Impl pops from _attributes/_dirty_attributes with default=None → safe, minimal.
2. QueryBuilder.table() — main calls Pivot.on(...).table(...) on a QueryBuilder; only DB/Schema/relationship had their own table, QueryBuilder did not. Sets _table, returns self. Correct.
3. QueryBuilder.without_global_scopes() — main calls it; undefined on main. Resets _global_scopes = {} (same shape as __init__). Correct.

Pivot refactor (core relationship code) — acceptable risk:

• Pivot storage moved from __original_attributes__ to _attributes[self._as] via Pivot() + set_raw_attributes(pivot_data, sync=True); __getattr__ reads _attributes, so product.pivot resolves.
• attach/detach/attach_related/detach_related now build pivot rows via connection.query().table(...).insert()/.where().delete(), and get_pivot_table_name receives builders (which expose get_table_name()) rather than models. The old path was non-functional on main anyway (it hit the 3 undefined methods). Full suite + new BTM integration tests pass.

Tests: real sqlite integration with real assertions; net-new (main had zero BelongsToMany/transaction coverage → no duplication vs #147/#154-#161). Transaction unit tests mock DB wiring appropriately (assert_awaited_once), not hollow.

Minor (non-blocking): three shallow assertions — assertIsNotNone(pivot) (L55, L63) and assertIsNotNone(builder) (L121). L121 is near-tautological (builder is never None). They still exercise the real code paths (which would raise if broken), so not a blocker; consider asserting on pivot contents / timestamp keys and on the where_exists clause in a follow-up.