Fix external-review issues: dedup OFF_MARKERS, paginate Bilibili, val…

[HarryXin0919]

…idate creators, version cap, retire bidao pilot, add tests + CI

Robustness / bugs:

• shared_markers.py: single source for OFF_MARKERS (FULL for features, COMPARE for compare_form -- the deliberate livestream / 联名 divergence preserved, behavior unchanged), ending the silent drift between the two copies.
• fetch_bilibili.py: paginate up to num (was first-page-only -> missed data past 50); add BilibiliError mirroring YouTubeError; factor out _to_record / _get_page.
• creators.py: validate_creators() lists ALL config problems (typo'd keys, missing alias/zone, duplicate alias, youtube missing channel) instead of a KeyError at fetch time; called at the start of fetch_multi.
• Cap bilibili-api-python <18.0.0 (requirements.txt + pyproject.toml).
• schema.py: VideoRecord TypedDict documenting the standard video record.

Cleanup -- retire the single-creator pilot cluster (bidao-hardcoded, mutually coupled via classified.json, superseded by the generalized features/scan_signals pipeline, and absent from every pipeline / README command list):

• delete classify_and_stats.py, fetch_videos.py, and the orphaned consumers comments.py
  • subtitle.py (they read the now-gone classified.json); drop all four from the PyInstaller spec hiddenimports; add shared_markers + schema.
• anonymize remaining bidao / real-BV references in analyze_video.py + diagnose.py docstrings; reword the README L2-text line to the surviving compare_meme.py.

Tests + CI (the project had zero tests):

• conftest.py + tests/ (23 tests): features.extract, off_tag, OFF_MARKERS divergence, validate_creators, spearman, build_report formatters, runtime paths, and fetch_bilibili parse/record/error + pagination (incl. the >50-video multi-page path).
• pyproject dev extra + pytest config; CI now runs pytest.

Deliberately skipped (with reasons): scripts/ -> subpackage restructure (would break the released PyInstaller packaging), web CSRF/rate-limit (localhost-only app), scipy for spearman (intentional zero-dep). Two report claims were verified already-fine and left alone: scan_signals already degrades on missing data; requirements/pyproject are already aligned.

Verified: 23 tests pass, compileall clean, viralens --help ok, frozen build succeeds and --vl-exec compare_form / fetch_multi / build_report all work (new modules bundled, BilibiliError surfaces correctly).