[Ilya Bogin (ilya-bogin-keenable)]

Follow-ups from the #368 review, one item per point.

1. Key in error text. KEENABLE_API_KEY is replaced with [redacted] in the HTTP error body, before the 1000-character cut, so a key straddling the cut can't leak a prefix either. Tests cover a key echoed in a 401 body (checked in fail_reason and conversation.json) and one at the cut.
2. Round deadline. time.monotonic() > deadline is checked again after each response, so a slow last query now fails the round too, as the docs say. Test: a single query whose response lands past the deadline.
3. Tests.
   • test_fresh_config_never_calls_keenable now has the model ask for search, then asserts the mock transport saw no request.
   • New tests cover a 302 (not followed, one request), a non-chat backend (refused before any request), and a non-JSON body (fails instead of returning empty results).
   • I mutated the code to check that each one fails on the regression you listed: follow_redirects=True, no backend check, invalid JSON read as [], no redaction, no post-response deadline check, and auto selecting Keenable (the old fresh-config test passed with that one).
4. Minor.
   • The config table now says search_max_num_results is per query for Keenable: 16 results, about 97k characters, per round with the defaults.
   • The data-boundary note says a query can repeat text from the document context or the skill as well.
   • A missing or non-string title falls back to the URL.

tests/test_officeqa_keenable_search.py: 42 passed. The full suite passes on 3.10 and 3.12, apart from failures that also happen on an unmodified ccfdcdf in my environment: test_backend_pi::test_default_model_from_env, two test_sleep_adopt_skill_subset tampered-manifest tests, and, on 3.10 only, test_officeqa_parallel_search::test_whole_mcp_request_has_a_deadline.