…croll

The guard `if timeout and timeout <= 0` short-circuits to falsy when
timeout=0, so the validation was skipped and the method fell through to
launching a real browser and navigating to the URL. In CI this caused the
unit test `test_ascrape_playwright_scroll_invalid_params` (which passes
timeout=0) to hang on a real network call until the 6h job timeout,
cancelling the entire Test Suite run.

Use an explicit `is not None` check so timeout=0 is correctly rejected.

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

…unit tests

With the 6h hang fixed, the Test Suite job now runs to completion and exposed
a pre-existing, broadly-broken suite: 100 failed / 208 passed. This restores it
to green (283 passed, 43 deselected) by scoping the unit job to genuine unit
tests and fixing real bugs uncovered along the way.

Scope CI to true unit tests:
- conftest: auto-mark tests/graphs/* (except the mocked abstract_graph_test) as
  `integration`. These run full pipelines against real LLM providers, remote
  search engines, or optional backends (OpenAI/Ollama/Mistral/Fireworks/Ernie),
  so they need API keys / network / extra packages and are not unit tests. The
  CI selector `-m "unit or not integration"` now deselects them.

Dependencies:
- add `ddgs` (langchain-community's DuckDuckGoSearchResults now imports it).
- add `selenium` and `pandas` to dev-deps so the mocked selenium-backend and
  CSV fetch unit tests can run.

Genuine source bug fixes:
- docloaders/chromium.py: close the browser in a `finally` (was leaked on
  failure); honor `requires_js_support` in `scrape()`; map the selenium backend
  to `ascrape_undetected_chromedriver` in lazy/alazy_load (was calling a
  nonexistent `ascrape_selenium`); validate browser_name before the retry loop;
  hoist the optional `Malenia` import to module scope.
- nodes/fetch_node.py: fix UnboundLocalError on `parsed_content`/`document`;
  default timeout to 30s when node_config is None (only explicit None disables);
  route `txt` input to the local-source handler.
- utils/copy.py: make `safe_deepcopy` a real recursive deep copy with a memo
  dict so circular references no longer raise DeepCopyError and object
  attributes are actually deep-copied.
- utils/proxy_rotation.py: parse scheme-less `host:port` servers correctly and
  route the `broker` keyword to proxy search; fix the error message.
- utils/research_web.py: slice DuckDuckGo results to max_results.
- docloaders/plasmate.py: hoist ChromiumLoader import to module scope.
- graphs/abstract_graph.py: raise a clear error when a model's provider cannot
  be determined (was a confusing KeyError).

Stale tests updated to assert current correct behavior (no assertions weakened):
chromium mocks, fetch fixtures, deepcopy/omni, generate_answer keys,
models_tokens limits, proxy/research_web/sys_dynamic_import, plasmate,
cleanup_html, robot/search nodes, script_creator_multi.

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