fix(lint): drop missing-page findings for pages that already exist

[chuenchen309]

Why

runSemanticLint passed the LLM's missing-page findings straight to the Review tab without validating them against the actual page set. The LLM doesn't reliably cross-reference the file list, so it flags entities whose page already exists — especially in non-English wikis, where its free-form titles (维特根斯坦, 阿德勒实体页缺失, 缺失页面：X) don't match a fixed prefix. The reporter saw 30 of 41 missing-page findings as false positives, drowning out the 11 genuine ones.

On the reported root-cause guesses in #537: the scan already covers entities/ (every wiki/*.md except log.md is summarized), so it isn't a "only scans concepts/" problem — the findings come from the LLM and simply weren't filtered against existing pages.

What

While summarizing pages, collect the set of existing page names (basename + frontmatter title, NFKC-normalized). Drop a missing-page finding when a known page name matches its title exactly or appears within it.

Fixes #537

Review notes

• Match strategy / over-match guard: exact match handles bare names; a substring fallback (guarded to names ≥ 2 chars) handles free-form/suffixed titles like 阿德勒实体页缺失. The substring pass could in principle drop a genuine missing-page whose title happens to contain a short existing page name. I chose the more permissive version because the reported problem is over-reporting — but if you'd prefer a stricter rule (exact-match only, or a higher length threshold), I'm glad to adjust.
• NFKC normalization folds full-width/compatibility forms so CJK/full-width variants compare equal.
• Only missing-page findings are filtered; other semantic types are untouched.

Tests

Added a regression test in lint.test.ts: an entity page exists but the (mocked) LLM reports it as missing-page with free-form Chinese titles — asserts the false positives are dropped and a genuinely-missing entity is kept. Fails before the change, passes after. Full unit suite: 1609 passed.

---

Disclosure: this change was produced by an AI agent (Claude Code) — root-cause analysis, fix, and regression test included. It's covered by the new test and the full unit suite passes; flagging for maintainer review.