Add consistency report and project glossary

[jserv]

A real-world deployment study [1] reported mainland-Chinese terms slipping past the linter in published zh-TW articles, blockquote citation contexts producing ~50 false positives across a 72-article corpus, and ASCII quotes auto-converted to 「」 inside YAML frontmatter breaking downstream parsers.

User-facing additions:

• '--consistency' reports mixed regional usage of one concept (both 線程 and 執行緒 in the same document). Groups by the rule's "english" anchor; skips TM-suppressed terms.
• '--exempt-blockquotes' (CLI + '[markdown]' config) excludes pulldown-cmark 'Tag::BlockQuote' ranges from scanning. Off by default: adopted blockquote prose is real content.
• YAML frontmatter preserves ASCII '"' / ''' scalar delimiters. Body prose still converts to 「」.
• '[glossary]' section in '.zhtw-mcp.toml': banned / preferred / proper_nouns lists. Banned terms inject synthetic Errors that TM cannot downgrade; proper_nouns suppress matching issues; both honor exclusion zones.
• Per-rule 'editorial_confidence' (low / medium / high) flows through issue inflation into MCP explain output. Low forces auto_fix_safe = false and needs_review = true. 優化, 算法, 場景 tagged low because both regional forms are valid zh-TW.

Calque-audit refinements:

• 消息 gains positional_clues; 好消息 / 壞消息 / 消息來源 no longer fire.
• Symmetric 元資料 rule mirrors 元數據 — both use to: [] plus english: "metadata", surfacing the English original as the preferred form. 詮釋資料 and 後設資料 (NAER terminology bank) remain unflagged as acceptable zh-TW alternatives.
• Real-world regression fixture pins the 14 documented blind-spot terms.

[1] https://ai-muninn.com/zh-TW/blog/zhtw-mcp-calque-blindspot-sweep

---

Summary by cubic

Adds a document-wide terminology consistency report and a project glossary to enforce preferred terms. Reduces false positives in blockquote citations and preserves ASCII quotes in YAML frontmatter.

• New Features

  • --consistency: groups by a rule’s english anchor to catch mixed regional terms in one doc; ignores TM‑suppressed issues.
  • Project glossary in .zhtw-mcp.toml ([glossary]): banned (always error), preferred (guides suggestions), proper_nouns (suppress); all honor exclusion zones.
  • Blockquote exemption: --exempt-blockquotes and [markdown].exempt_blockquotes = true exclude Markdown blockquotes from scanning (off by default).
  • Per‑rule editorial_confidence (low/medium/high): included in diagnostics; low forces auto_fix_safe = false and needs_review = true.

• Bug Fixes

  • YAML frontmatter now preserves ASCII " and ' scalar delimiters; body prose still converts to 「」.
  • Calque rules: added positional clues for 消息 (no more false positives like 好消息/壞消息/消息來源); added a symmetric 元資料 rule mirroring 元數據 with english: "metadata" while keeping 詮釋資料/後設資料 acceptable.

^{Written for commit 269c8dd. Summary will update on new commits.}

[cubic-dev-ai]

7 issues found across 29 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/rules/glossary.rs">

<violation number="1" location="src/rules/glossary.rs:86">
P1: Skipping synthetic banned issues when a regular issue already exists can let TM downgrade the only report, breaking the intended `banned > TM` precedence.</violation>
</file>

<file name="src/engine/consistency.rs">

<violation number="1" location="src/engine/consistency.rs:148">
P1: Do not bypass group matching when only one term group exists; it can select unrelated glossary terms and create false consistency diagnostics.</violation>
</file>

<file name="src/mcp/tools.rs">

<violation number="1" location="src/mcp/tools.rs:853">
P1: Fix-mode ordering lets TM downgrade glossary-banned synthetic errors, breaking the intended `banned > TM` precedence.</violation>

<violation number="2" location="src/mcp/tools.rs:1083">
P2: `tools/list` schema was not updated for the new `exempt_blockquotes`/`glossary`/`consistency` arguments, causing API contract drift.</violation>
</file>

<file name="src/main.rs">

<violation number="1" location="src/main.rs:869">
P2: The new `--exempt-blockquotes` mode is not represented in scan-cache keys, so cached results can be incorrect when toggling the flag.</violation>

<violation number="2" location="src/main.rs:1194">
P1: Cache-hit paths can drop source text needed by the new glossary/consistency features, causing missed findings.</violation>
</file>

<file name="tests/realworld_calques.rs">

<violation number="1" location="tests/realworld_calques.rs:62">
P2: Match on containment here so the collocation regression is caught even when the scanner reports the full phrase instead of the bare term.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P2: The new --exempt-blockquotes mode is not represented in scan-cache keys, so cached results can be incorrect when toggling the flag.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/main.rs, line 869:

<comment>The new `--exempt-blockquotes` mode is not represented in scan-cache keys, so cached results can be incorrect when toggling the flag.</comment>

<file context>
@@ -825,6 +866,9 @@ fn run_lint_batch(params: &LintBatchParams<'_>) -> Result<()> {
     if params.relaxed {
         cfg = cfg.with_relaxed();
     }
+    if params.exempt_blockquotes {
+        cfg = cfg.with_exempt_blockquotes(true);
+    }
</file context>