feat(web_search): SearXNG meta-search with DuckDuckGo fallback

[netty-linux]

Summary

Adds an optional SearXNG backend for web_search — self-hosted meta-search, aggregates multiple upstream engines rather than depending on any single one's bot-detection posture. Tried first when configured via [web_fetch].searxng_base_url (or the app-side SEARXNG_BASE_URL env var, once threaded through); any failure (unreachable, timeout, non-200, malformed JSON, JSON format disabled on the instance) falls through to the existing DuckDuckGo HTML scraping — a misconfigured or down self-hosted instance must never make search unavailable when a working zero-infra fallback exists.

Increment 1 of 4 in the approved plan (SearXNG → Crawl4AI/Wayback → Postgres batching → OTLP+StateSummary).

Changes

• nullain_tools.web_search: _searxng_search() queries {base_url}/search?format=json, returns None (never raises) on any failure signal. create_web_search_tool(searxng_base_url=...) tries it first, falls through to the unmodified DuckDuckGo parsing path.
• nullain.config.WebFetchConfig: new searxng_base_url field. None (default) preserves exact prior behavior — no attempt to reach any SearXNG endpoint at all.
• register_default_tools/Agent facade: threaded through the same way web_fetch_headers/web_search_headers already are.

Infra note (not part of this PR)

Operating a SearXNG instance (+ typically Redis/Valkey for its own rate limiter) is the deploying operator's responsibility — this change only knows how to consume a configured instance's JSON API. On Coolify: add SearXNG as a separate Resource on the same internal network, point SEARXNG_BASE_URL at it.

Test plan

• 10 new tests: SearXNG success skips DDG entirely; SearXNG failure modes (unreachable, HTTP error, malformed JSON, empty results) all fall through to DDG; default (searxng_base_url=None) never attempts to reach any SearXNG endpoint.
• All 10 pre-existing test_web_search.py tests pass unmodified — confirms the DDG path itself wasn't touched by the refactor.
• Full suite: 658 passed, 7 skipped, 84.75% coverage.
• uv run ruff check . / uv run ruff format --check . / uv run pyright — all clean (same invocations CI uses).