v2.16.0 — The Crawler That Refuses to Die

The free-proxy rotation chain: harvest from public lists → validate → rotate
per request → feed every outcome back → refill on exhaustion. The engine
v2 and the self-healing freshness layer already existed; this change wires
them into the crawl path, fixes their deployment blockers, and adds the
first tests they ever had.

Fixed

• proxy_engine_v2: cache path was a hardcoded absolute path from another
  machine (/home/yu/...) — now project-relative .crawl_cache (env
  OMK_CACHE_DIR still wins). IPv6 probe no longer runs on the constructor's
  critical path (lazy, once). Burned proxies are now persisted (in-memory
  burn died on restart and the cache resurrected known-bad endpoints).
• Protocol detection read the URL scheme, so every https-served text list
  (all of GitHub raw) was labeled https. Detection now uses filename/query
  markers (http.txt, protocol=socks5, type=http, HTTPS_RAW,
  /socks5_list/) with the scheme excluded.
• aiohttp/aiohttp-socks were lazy-imported but declared nowhere — new
  proxy extra (pip install 'omk-crawl[proxy]'), included in all.

Added

• Freshness ("newest"): ProxyNode.harvested_at/source; a source
  re-serving a proxy touches its freshness instead of duplicating it;
  STALE_SECONDS (default 1h, OMK_PROXY_STALE_SECONDS) demotes stale
  entries in scoring and next(fresh_only=True) excludes them.
• rotate_crawl.py — RotationCrawler: one pool proxy per attempt over
  curl_cffi with the node's assigned TLS profile; outcome feedback
  (ok→score up, 429→per-proxy backoff, 407→burn, blocked/error→failure
  streaks); bounded synchronous refill when the pool empties mid-crawl;
  crawl_many() bounded concurrency. pool_cycles() is the bounded
  무한의 굴레 loop (cycles=0 → until cancelled).
• proxy:// scheme + ProxyPoolTool: proxy://stats (offline),
  proxy://refresh, proxy://loop?cycles=N&interval=S. Bad query params
  fail closed as results, never raise (BaseTool contract).
• CLI --proxy-pool: crawl a URL through the rotation pool.
• 30 offline tests (tests/test_proxy_rotation.py); suite now 714 passing.

Verification boundaries

• Live harvesting/validation against public proxy sources and the Firebase
  validation target are network paths exercised only by the live lane — the
  offline suite covers parsing, scoring, selection, feedback, persistence
  and CLI plumbing with fakes.
• Same-domain-style caveat: free proxies are untrusted third parties —
  credentials or personal data must not transit them; the pool is for
  public-content crawling only.