fix(sentry): drop expected upstream provider errors

[sjawhar]

Problem

With Sentry enabled, every provider rate-limit and overload response becomes an unhandled Sentry error. In our production deployment RateLimitError: Error code: 429 accumulated 56 unhandled events in three days, alongside 529 overloaded_error, and it sits at the top of the issue stream above defects that are actually ours.

These are not proxy defects. A 429 is the backend applying backpressure, and the pipeline already handles it deliberately: _handle_anthropic_error converts AnthropicStatusError into a BackendAPIError, backend_api_error_handler returns it to the client in Anthropic error format, and the caller retries.

Root cause

The Sentry SDK's Anthropic integration is auto-enabled and captures at the SDK call site, so the event is created inside anthropic/_base_client.py:request with mechanism: {"type": "anthropic", "handled": false} before our handler ever sees the exception. Nothing downstream can un-capture it.

Fix

_sentry_before_send drops APIStatusError whose status_code is one of 408, 429, 500, 502, 503, 504, 529 — throttling and availability, the cases the pipeline converts and the caller retries. The check is on the status code rather than the exception class so that an SDK renaming or adding a status subclass cannot silently start reporting again.

Verification

• Four new tests in test_sentry_scrubbing.py build real APIStatusErrors from httpx responses: 429 and 529 are dropped, 400 is kept, and an ordinary TypeError from our own code is kept. The two drop tests fail before this change and pass after; the two keep tests pass throughout, so they pin the behavior that must not regress.
• uv run pytest tests/luthien_proxy/unit_tests/test_sentry_scrubbing.py passes at this commit (49 passed).
• dev_checks.sh: shellcheck, ruff format, ruff lint, ruff docstrings, and pyright all clean. The pytest stage shows the same three machine-local failures documented in fix(sentry): never capture opentelemetry.context detach noise #807 and fix(pipeline): serialize stream events with model_dump(mode="json") #808 (test_onboard.py::test_find_docker_ports_respects_env_vars, two in test_config_registry.py); all three pass when run with GATEWAY_PORT/POSTGRES_PORT/REDIS_PORT unset, so they are this box's environment rather than this change.

---

Update: also dropping 400/404/401

This PR originally kept 4xx codes other than 429 on the theory that a 400 means an actionable malformed request. Three more Sentry issues showed that isn't the whole picture — the pipeline already treats every AnthropicStatusError identically regardless of status code (_handle_anthropic_error / _build_error_event both log at warning and convert to BackendAPIError for any status code), and these three are the provider legitimately rejecting something the client sent, not something the proxy built:

• LUTHIEN-6 (1,080 events, by far the largest single issue in the whole quota drain): a client sent message content with a field Anthropic's schema rejects (400, e.g. messages.297.content.1.text.parsed_output: Extra inputs are not permitted). Confirmed parsed_output is not a field this proxy or any policy injects — it's client-originated.
• LUTHIEN-2 (125 events): a client requested a model name Anthropic doesn't have (404, e.g. model: opus-4.6).
• LUTHIEN-D (28 events): a client-supplied bearer token that Anthropic itself rejected as invalid (401) — passthrough-credential mode forwards the client's own token unchanged; the proxy didn't mint it.

_EXPECTED_UPSTREAM_STATUS_CODES now includes 400, 401, 404 alongside the original throttling/availability codes. Three new drop-tests (test_drops_upstream_bad_request_error, _not_found_error, _authentication_error) plus a new keep-test for a status code outside the set (test_keeps_upstream_status_code_outside_expected_set, using 403) so a genuinely new upstream failure mode stays visible.

Silences: LUTHIEN-6, LUTHIEN-2, LUTHIEN-D (new), plus the original LUTHIEN-8/rate-limit and overload noise this PR already covered.