fix(weekly): restore voice pointer and zh-CN companion

[atomchung]

Refs #676
Refs #683

User before / after

Before, SKILL.md reintroduced a route-local copy of the universal decision-first voice rule, and a zh-CN weekly review fell through to English companion prose. After, the runtime routes the considered-trade structure through #676's single global output-voice authority, and the weekly companion renders entirely in Simplified Chinese for zh-CN.

Scope

• Restore docs: add global output voice contract #700's pointer-only wording in SKILL.md, retaining the weekly-market-read CLI name.
• Add a zh-CN WeeklyMarketRead renderer, including its visible NextWeekWatch line.
• Add regressions for the retired route-local fragment and a real prepare -> preview -> weekly-market-read zh-CN lifecycle.

No L1 lookup, memory, scheduler, persistence, VIX threshold, metrics, diagnosis, or canonical-state change. PR #703 is provenance-only; PR #661 is a separate, stale rationale lane. Neither overlaps this fix.

Evidence

• Focused: python3.12 tests/test_output_voice.py and python3.12 tests/test_weekly_market_read.py pass.
• Full offline: python3.12 tests/run_all.py passes.
• The lifecycle witness runs the real CLI path using fictional artifacts, requires preview before the read, asserts Simplified headings and no English companion fallback, and keeps the existing no-provider-fetch regression.

QA boundary

This is reproducible agent-simulated CLI/lifecycle evidence, not owner-live UX acceptance. The formal weekly walkthrough must run after this repair is on latest origin/main, with the owner completing the receipt and verdict. No merge is requested by this PR.

[atomchung]

Post-merge re-review — the two intended fixes are correct; one locale regression remains

Reviewed the merged result on main@8f6d390 and the final PR-head CI.

The #704 changes themselves are sound:

• SKILL.md is back to #676's pointer-only global voice authority, and the new negative regression prevents the retired route-local sentence from returning;
• zh-CN now has an explicit renderer and a real prepare → preview → weekly-market-read lifecycle witness;
• PR-head workflow run #1349 passed the full offline suite on Python 3.11 and 3.12.

I found one remaining user-visible locale defect in the surrounding #702 code:

zh-TW still renders the NextWeekWatch line in English

render_zh_tw() writes watch['check'] and watch['trigger']['condition'] directly. Both values are created by _watch() in English:

• Before explaining a move with the market...
• Before treating a lower price as evidence...
• VIX rose in the frozen review window

So a Traditional-Chinese weekly review currently produces Chinese headings and analysis followed by an English 下週關注 line. #704 fixed the equivalent zh-CN path by introducing localized watch_checks, but left the original zh-TW path unchanged.

The current tests do not catch this: the zh-TW assertion checks headings and absence of engine fact / agent judgment, not the English watch/trigger strings. The new zh-CN lifecycle also exercises only the first read, not every offered focus branch.

Smallest repair

Keep this under #683; no new issue or architecture is needed.

• localize the zh-TW watch check and trigger instead of rendering the English domain strings;
• add a focused regression covering business_evidence, position_size, and the skipped/default watch for both zh-TW and zh-CN;
• run one zh-TW real CLI lifecycle with the second --focus read and assert the companion contains no English watch fragments;
• no change to WeeklyMarketRead, trigger logic, persistence, provider calls, or #676.

After that narrow hotfix, I see no further code blocker to a private owner weekly-review walkthrough. The VIX delta > 0 calibration remains an owner-live product question, not a reason to widen this patch.