[lesnik512]

Closes #310. Closes #340.

This PR changes two Sentry defaults that cost time on every request or log record:

field	was	now
sentry_auto_session_tracking	True	False
sentry_logging_breadcrumb_level	logging.INFO	logging.WARNING

_warn_breadcrumb_level_ignored used to compare against logging.INFO to tell whether the user had changed the field. It now compares against logging.WARNING, so an explicit INFO is treated as a change from the default. When sentry_integrations brings its own LoggingIntegration, or sentry_default_integrations=False is set, that explicit INFO is ignored and lite-bootstrap warns.

Tests

Written first, all failing on main:

• test_sentry_does_not_track_sessions_by_default: the client's auto_session_tracking option is False.
• test_sentry_breadcrumbs_start_at_warning_by_default: the breadcrumb handler's level is WARNING.
• test_sentry_error_event_carries_no_info_breadcrumbs_by_default: an INFO and a WARNING record logged before an error leave only the WARNING breadcrumb on the event. The test uses its own logger and filters by category, because the event also carries a subprocess breadcrumb from sentry-sdk working out the release.
• test_sentry_warns_when_a_non_default_info_breadcrumb_level_is_ignored: on main, INFO was the default, so no warning was raised.

test_sentry_logging_breadcrumb_level_controls_the_breadcrumb_handler gains a WARNING case. The full suite passes at 100% coverage on sentry-sdk 2.73.0 and 3.0.0b2, and ty passes against both. just lint, just adr-check, just docs-build and lychee pass.

Measurement

#340 asks for a rerun showing that an INFO call at WARNING costs about what it costs at None. Each mode ran in its own process: one structlog log.info call through FreeBootstrapper, output to /dev/null, median of 5 × 20000 calls, measured in-process on an M2 Mac.

sentry-sdk	no Sentry	breadcrumbs at INFO	at WARNING	None
2.73.0	10.04 µs	15.35 µs	10.26 µs	10.29 µs
3.0.0b2	9.55 µs	15.75 µs	10.56 µs	10.50 µs

The new default saves about 5 µs per INFO record on both majors. I didn't rerun the session cost for #310. The 7.7 µs per request in the docs and the 7.2 to 8.7 µs re-measured in #310 both come from the issue, and the default change affects nothing else.

Docs

• configuration.md: the two bullets give the new defaults and say how to get the old behaviour back. The paragraph saying breadcrumbs "stay on by default" is rewritten.
• performance.md:
  • The session row is marked as the default since 2.0.
  • The "applies one saving for you" paragraph now lists all three savings, along with what the two new ones cost you.
  • It says the page's numbers were measured against the pre-2.0 defaults.
• benchmarks/README.md:
  • The tuned snippet, which performance.md includes, drops sentry_auto_session_tracking=False because that is now the default.
  • A line under the snippet records that the tuned runs set it explicitly.
• sentry_scenarios.py: the errors_only comment says it reproduces 1.x's defaults. Its kwargs are unchanged, so the numbers measured with it are still valid.

Breaking changes

• Release health: crash-free session and user rates stop. Set sentry_auto_session_tracking=True to keep them. Only the HTTP frameworks opened sessions per request, so FastStream and Free behave the same as before.
• Breadcrumbs: error events no longer carry INFO log breadcrumbs. Set sentry_logging_breadcrumb_level=logging.INFO to keep them.