Hardening pass: scheduler backoff, FK cascade, a11y, test safety net

Pre-traffic cleanup from an audit:

* Scheduler: poll_due_sources now keys on the last *attempt* (success or
  failure), not the last success, and scales the wait by the consecutive-
  failure streak (capped at a day). A failing feed (e.g. Phys.org's HTTP 429s)
  used to be retried every cycle because it had no successful run; it now backs
  off and recovers on its own. Extracted due_source_rows() + tests.

* FK hygiene: deleting a daily_brief is supposed to cascade to its items, but
  SQLite enforces foreign keys per-connection — connect() already sets the
  pragma, so the cascade is correct going forward; added a regression test.
  (Orphaned items + Phys.org settings were cleaned directly on the live DB.)

* a11y: modal/drawer dialogs are now focusable (tabindex), close on Escape
  (window) and on backdrop click via a target check (dropping the inner
  stopPropagation handlers). Build is warning-free.

* tests: conftest points any un-mocked LLM client at a closed port with a 1s
  timeout, so an accidental real call fails fast instead of hanging the suite.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/conftest.py

@@ -0,0 +1,16 @@
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def _no_real_llm(monkeypatch):
+    """Safety net: no test should reach the real arbiter.
+
+    If a test constructs an LLM client from env without mocking, point it at a
+    closed local port with a 1s timeout so the call fails fast instead of
+    hanging on the production 300s timeout (which looks like a stalled suite).
+    Tests that exercise the LLM patch it at the function level, which overrides
+    these env defaults.
+    """
+    monkeypatch.setenv("GOODNEWS_LLM_BASE_URL", "http://127.0.0.1:1/v1")
+    monkeypatch.setenv("GOODNEWS_LLM_TIMEOUT", "1")
+    monkeypatch.delenv("GOODNEWS_LLM_API_KEY", raising=False)