Stop configuring logging at import time

[gitronald]

Implements .planners/plans/057-logging-no-import-side-effect/plan.md

Removes the module-scope Logger().start(__name__) calls (ten sites) so plain import WebSearcher no longer attaches a JSONL handler to the root logger or forces root to DEBUG. Modules now use plain logging.getLogger(__name__); the package installs a NullHandler on its own logger; crawl-time configuration in SearchEngine.__init__ is untouched.

Verified:

• import WebSearcher leaves logging.getLogger().handlers == [] (was a root StreamHandler + level DEBUG)
• a downstream logging.basicConfig(...) after import now takes effect (was a silent no-op)
• constructing a requests-backend SearchEngine still installs the JSONL console + file sinks and writes the crawl-log schema unchanged
• 643 tests and 102 snapshots pass; ruff/pyrefly clean

[gitronald]

Code review (xhigh: 10 angles → verify → sweep)

Ten independent finder angles (line-scan, removed-behavior audit, cross-file trace, language pitfalls, logger-hierarchy correctness, reuse, simplification, efficiency, altitude, conventions), one adversarial verifier per deduped candidate, then a fresh-eyes gap sweep. Five candidates survived; verdicts and outcomes:

1. CONFIRMED, fixed (b3fd8df) — Logger.start() kept a dead, misleading default (name bound to logger.py's own __name__, so an argless call would silently yield a logger named WebSearcher.logger), and nothing warned that start() runs dictConfig and reconfigures root process-wide. name is now required and the docstring states the entry-point-only constraint.
2. CONFIRMED, fixed (b3fd8df) — nothing pinned the PR's core invariant in CI. Added test_import_configures_no_logging: a subprocess probe (fresh interpreter — in-process assertions would be order-dependent, since other tests run dictConfig and WebSearcher is already in sys.modules) asserting root handlers stay [], root level stays WARNING, and the package NullHandler is present.
3. CONFIRMED, fixed (b3fd8df) — the changelog entry omitted a second observable change: import no longer force-sets third-party logger levels (requests/urllib3 → WARNING, asyncio/chardet.charsetprober/parso → INFO), so an app running root at DEBUG now sees urllib3 chatter from the SearchEngine-free HTTP helpers (e.g. download_locations). Clause added.
4. CONFIRMED behavior, conscious no-op — the NullHandler addHandler is not reload-idempotent (importlib.reload stacks handlers). Kept unguarded deliberately: requests and urllib3 ship the identical unguarded idiom (verified in this venv's installed sources), NullHandler.emit is a no-op so stacking has no functional effect, and %autoreload only re-executes __init__.py when that file itself changes.
5. Conscious no-op — no ruff banned-api (TID251) guard confining Logger().start to searchers/. The new regression test pins the invariant in CI; a lint ban would add config surface for marginal gain over that.

Verified clean: no consumer anywhere reads the four deleted module log attributes; logger names are identical pre/post change (probed); no propagate=False orphaning; demos/CLI use print() for UX; README's logging example is crawl-path only (still accurate); no other import-time logging mutation remains in the package.

Gate after fixes: 644 tests + 102 snapshots pass, ruff + ruff-format + pyrefly clean. Gap sweep over the post-fix diff: no new findings.