test(logger): remove redundant tests and parametrize duplicates in test_logger.py (#830)

## Summary

Cleans up `py/src/braintrust/test_logger.py` by removing redundant and
near-empty tests and parametrizing near-duplicates. Test-only change; no
SDK code is modified.

- **4,999 → 3,436 lines**, **223 → 189 tests** in `test_logger.py` (one
test moved to `test_gitutil.py`).
- Each commit is one self-contained batch and can be reviewed on its
own.

### What changed

- **Removed no-op or duplicated tests**, for example:
  - `init` dataset-id tests that never called `init()`
  - `JSONAttachment` tests that only checked objects they had just built
  - `span.log` binary/bad-key tests that asserted only that a key exists
- attachment, deep-copy, `id_gen` and `span_components` tests already
covered in `test_bt_json.py`, `test_id_gen.py` or
`test_span_components.py`
- **Parametrized near-duplicates:**
  - generator `MAX_GENERATOR_ITEMS` limits (sync/async × 3/0/-1)
  - span export format (9 tests → 1)
  - `model_dump` metadata across 7 logging APIs
- masking across logger, experiment, child span and dataset: one shared
masking function replaces 4 hand-written ones
  - prompt/parameters version-over-environment
  - strict-mode templates
  - bt-eval BTQL merge
  - `atexit`, `proxy_conn`
  - `emit_log` `str.format`/t-string template rendering
- **Dissolved the `TestLogger` grab-bag** into plain pytest functions.
- **Replaced manual `os.environ` save/restore** (and a hand-built
`pytest.MonkeyPatch()`) with the `monkeypatch` fixture.
- **Moved** `test_get_repo_info_without_settings_returns_none` to
`test_gitutil.py`.

### Assertions that got stricter

- Generator limit 0 now also checks that no truncation warning is
logged. The old tests missed a regression there.
- The inferred child-span name is now checked exactly
(`user_fn:user_module.py:2`) by calling `start_span()` from a module
that isn't under `braintrust.*`. Before, the test only checked that the
name wasn't empty.
- The experiment span-link test checks the full URL instead of three
substrings.
- Template-rendering cases check the full metadata, the log level, and
that no `error` field is set.
- The masking test checks that no unmasked value appears anywhere in the
logged rows.
- New case: `BRAINTRUST_OTEL_COMPAT` takes priority over
`BRAINTRUST_LEGACY_IDS` for the export format.

## Test plan

- [x] `nox -s test_core`: 926 passed, 69 skipped, 12 xfailed (main: 955
passed). Outside `test_logger.py`, the only difference is the test moved
into `test_gitutil.py`.
- [x] `test_logger.py` passes in 5 shuffled orders, and on Python 3.11
(t-string cases skip) as well as 3.14.
- [x] Mutation checks: for each rewritten area (generator limits, export
selection, masking, version precedence, BTQL merge, template rendering,
inferred span names), I broke the code under test in `logger.py` and
confirmed the new tests fail.
- [x] `ruff`, `pre-commit` hooks, `pylint --errors-only`.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Author: AbhiPrasad

py/src/braintrust/test_gitutil.py

@@ -63,3 +63,8 @@ def test_repo_info_returns_none_when_git_cannot_be_resolved(monkeypatch: pytest.
     assert gitutil._current_repo() is None
     assert gitutil.repo_info() is None
     assert list(gitutil.get_past_n_ancestors()) == []
+
+
+def test_get_repo_info_without_settings_returns_none():
+    """Direct call to get_repo_info with settings=None should return None."""
+    assert gitutil.get_repo_info(None) is None