feat(dataflows): CLS / Cninfo / Eastmoney news for HK and CN A-share tickers

[chitboon]

Summary

Adds first-class news coverage for HK and CN A-share (SH / SZ) tickers
via three new vendors plus a ticker-suffix auto-router. Purely additive at the
data layer — US tickers are unchanged.

Suffix	Vendors
.HK	Eastmoney + CLS
.SS / .SH / .SZ	Eastmoney + CLS + Cninfo
else (incl. US)	yfinance (existing behaviour)

The router engages when data_vendors.news_data is set to "auto". This PR
flips the default from "yfinance" to "auto" so coverage is on out of the
box; setting it back to "yfinance" (or "alpha_vantage") restores legacy
single-vendor behaviour.

What's in this PR (3 commits)

1. Per-vendor cache scaffolding — daily / weekly bucket cache module used
   by the new vendors. Reduces repeat-fetch cost when analysing multiple
   tickers in the same day or re-running the same ticker the same week.
2. yfinance ticker normaliser (.SH → .SS) — Yahoo only accepts .SS
   for Shanghai; users pasting .SH from other sources now route correctly.
   Also tightens get_global_news_yfinance from 4 overlapping queries to 2,
   roughly halving macro-news fetch latency.
3. CLS / Cninfo / Eastmoney vendors + auto-router — vendor modules,
   route_to_vendor multi-source merge, default_config flip, akshare>=1.18
   dep, tests.

What's NOT in this PR (held back deliberately)

• SGX (Singapore Business Times RSS) vendor — needs more end-to-end
  testing before shipping; will follow as a separate PR.
• Analyst-prefetch refactor — the consuming news_analyst.py already
  routes through get_news → route_to_vendor, so the new vendors engage
  without analyst-side changes. The single-shot rewrite is a separate
  architectural concern and not part of this PR.

Failure modes

• A per-vendor failure or [Vendor skip — ...] marker is dropped from the
  multi-source merge; the analyst sees whichever sources succeeded.
• If all sources fail / skip, the router falls through to yfinance as a final
  safety net — never raises.

Tests

• 4 new test files; all use unittest.mock at the akshare import boundary
  — no live network calls in CI.
• Covers: routing per suffix, multi-vendor merge, skip-marker filtering,
  cache hit/miss across daily and weekly buckets, Cninfo's two known schema
  shapes (公告链接 direct vs announcementId + orgId reconstruction).
• Full suite remains fast (~5s).

New dep

• akshare>=1.18 — wraps Eastmoney + Cninfo endpoints. Imported lazily
  inside the vendor modules; users analysing only US tickers pay no import
  cost.

Test plan

• python -m pytest tests/ — all green locally
• _resolve_auto_vendors smoke covers .HK, .SS, .SH, .SZ, .SI,
  AAPL, and get_global_news
• Reviewer to confirm akshare install lands cleanly in CI
• Reviewer to confirm the news_data: "auto" default flip is desired —
  happy to revert to "yfinance" and document "auto" opt-in if preferred

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces comprehensive support for Chinese-market data (HK, Shanghai, and Shenzhen) by integrating the akshare library and adding new news vendors: Eastmoney, CLS, and Cninfo. It implements an "auto" routing mechanism that selects the appropriate vendor based on ticker suffixes and merges results from multiple sources for Asian markets. Additionally, a new on-disk caching layer is introduced to optimize performance for slow-changing data like financial statements and global news. Feedback was provided regarding the caching mechanism to ensure atomic file writes, preventing potential data corruption during concurrent access.

[gemini-code-assist]

Writing directly to the cache file can lead to corrupted data if the process is interrupted or if multiple processes attempt to write to the same file simultaneously. It is safer to write to a temporary file first and then use os.replace() to perform an atomic update.

Suggested change

	with open(_path_for(key), "w", encoding="utf-8") as f:
	f.write(value)
	temp_path = f"{_path_for(key)}.tmp"
	with open(temp_path, "w", encoding="utf-8") as f:
	f.write(value)
	os.replace(temp_path, _path_for(key))

[chitboon]

Addressed in f939f5f — _write now uses tmp + os.replace for atomic publish.

[chitboon]

Thanks for the catch — addressed in f939f5f. _write now streams to <path>.tmp and uses os.replace to publish atomically, so a crash mid-write leaves the prior cache file intact and concurrent writers race only on the rename.