update parsers and classifiers for v0.6.7a2 demo data

[gitronald]

Summary

Fix 5 classification issues and 3 parser issues found in data/demo-ws-v0.6.7a2/ demo data. All were producing unknown types, parse errors, or null fields — every change is a net improvement with no working behavior regressed.

Classifier fixes

• Tag-agnostic heading search — headings changed from <h2>/<div> to <span> with aria-level + role="heading". Replaced two div-specific searches with one tag-agnostic find_all(attrs=...). Fixes 3 misclassifications: people_also_ask, perspectives, searches_related.
• discussions_and_forums text check — classifier matched any div.IFnjPb[role=heading] regardless of text. Now requires heading text to start with "Discussions and forums".
• "What people are saying" mapping — moved from knowledge to perspectives.

Parser fixes

• general.py — extracted find_subcomponents() to separate discovery from parsing. Added PmEWq video format so video results no longer fall through to the wrapper.
• knowledge.py — deduplicated URLs in AI overview (24 links with repeats down to ~13 unique).
• top_stories.py — added get_title() helper with multi-selector fallback for perspectives titles (eAaXgc in addition to n0jPhd).

Tests

• Rewrote test_parse_serp.py with syrupy snapshot test (parametrized by serp_id) + 8 structural validation tests. All 9 pass.

Tooling

• Added scripts/demo_screenshot.py for visual SERP inspection with component type highlights injected via BeautifulSoup.

Notes for future parser updates

• Use scripts/demo_screenshot.py to visually inspect component boundaries and classification before diving into code.
• Heading tags change frequently — the tag-agnostic attrs={"aria-level": ..., "role": "heading"} pattern should keep working, but watch for attribute changes too.
• Classifier order matters — ClassifyMain.classify() runs an ordered chain; first non-"unknown" wins. Check the full chain when adding or modifying classifiers.
• Multi-selector helpers (like get_title(), get_cite()) are the preferred pattern for handling CSS class changes — add new selectors to the list rather than replacing old ones.
• find_subcomponents() is the single place to add new general result formats.
• Run .claude/commands/parser-update.md for the full 7-phase diagnostic workflow.

Test plan

• poetry run pytest tests/test_parse_serp.py -v — all 9 tests pass
• Visual inspection via demo_screenshot.py confirms correct component boundaries
• No regressions — all previously correct classifications unchanged