Summary

PR1 of the 2026-06-05 bug-audit-v2 plan. Closes ten lifecycle / teardown findings — every fix is "bootstrap touched state, teardown didn't clean it up" or "constructing two bootstrappers around the same app caused stacking." Sequenced per planning/specs/2026-06-05-bug-audit-v2-sequencing.md; per-task TDD plan at planning/plans/2026-06-05-pr1-lifecycle.md; parent audit at planning/specs/2026-06-05-bug-audit-v2.md.

Findings closed

• LOG-1 — Document OTel set_tracer_provider set-once constraint. The SDK enforces it via _TRACER_PROVIDER_SET_ONCE.do_once(...); the "global reset on teardown" fix in the audit isn't achievable without touching SDK internals, so this becomes a class docstring that names the constraint and the supported lifecycle.
• LOG-2 — Restore the two OTel-namespace stdlib loggers (opentelemetry.instrumentation.instrumentor, opentelemetry.trace) to their pre-bootstrap disabled state on teardown. Capture-and-restore via a new _prior_logger_disabled: dict[str, bool] field.
• LOG-3 — LoggingInstrument.teardown() no longer dies on a handler.close() raise mid-loop. Errors are aggregated into TeardownError (matching BaseBootstrapper.teardown's pattern) so the rest of teardown — level reset, factory close — always runs. Factory-close errors are also caught and included in the aggregate so they don't silently replace handler errors via finally-replaces-pending semantics.
• LOG-4 — FastStreamLoggingInstrument captures broker.config.logger.params_storage in bootstrap and restores it in teardown. Symmetric to LOG-2 on the broker side.
• LOG-5 / SEC-4 — Replace assert precondition checks with explicit raises in _narrow_app (TypeError, per ruff's TRY004) and PyroscopeInstrument.bootstrap (RuntimeError, since the check is is None not isinstance). Closes both bandit B101 findings.
• LOG-6 — Replace LitestarOpenTelemetryInstrumentationMiddleware._otel_apps from dict[int, ASGIApp] (id-keyed, never evicts) to weakref.WeakKeyDictionary[ASGIApp, ASGIApp] so wrapper apps GC with their next_app. Both .get() and __setitem__ are wrapped in contextlib.suppress(TypeError) so non-weakrefable callables fall through unharmed.
• LOG-7 — FastMcpBootstrapper.__init__ now inspects application.providers for an existing _TeardownProvider and warns + skips re-attachment, preventing stacked teardown invocations when two bootstrappers are constructed around the same FastMCP app.
• LOG-8 — FastAPIBootstrapper.__init__ uses application._lite_bootstrap_lifespan_attached (private attribute on the FastAPI instance) as a sentinel to detect and warn on duplicate lifespan-wrap attempts. Uses the library-private-attribute convention rather than squatting in Starlette's user-facing application.state namespace.
• LOG-9 — SentryInstrument.teardown() calls sentry_sdk.flush(timeout=2) then sentry_sdk.init() (no args) to disable further event capture. The two existing tests that called sentry_sdk.init() in finally as a workaround now use instrument.teardown() instead.

Implementation notes

• Two reviewer-driven course corrections during execution:
  • LOG-3 first landed with raise close_errors[0] (lost info on multiple failures and silent factory-close-error masking); rewritten to use the project's existing TeardownError(errors) pattern.
  • LOG-6's first landing only wrapped __setitem__ in contextlib.suppress; WeakKeyDictionary.get() also raises TypeError on non-weakrefable keys, so both call sites were wrapped and the test was rewritten to use a real non-weakrefable class (__slots__ = (), no __weakref__ slot).
  • LOG-8 first stored the sentinel on application.state; moved to a _lite_bootstrap_lifespan_attached direct attribute matching opentelemetry-instrumentation-fastapi's approach.
• Two non-blocking follow-ups flagged by the final reviewer for PR2/PR3 scope:
  • Lift TeardownError aggregation pattern from LoggingInstrument to other instruments' teardown methods (currently only LoggingInstrument and BaseBootstrapper use it).
  • Document the "one bootstrapper per application" + "one OpenTelemetryInstrument per process" lifecycle constraints in README.md.

Test plan

• just test — 164 passed (10 new tests + 1 updated existing), 100% coverage.
• just lint-ci — clean (ruff format, ruff check, eof-fixer, ty).
• bandit -r lite_bootstrap — no issues (both prior B101 findings closed by LOG-5/SEC-4).
• Each commit is scoped to one finding ID so review/revert can land at the task level.

🤖 Generated with Claude Code