[argszero]

Summary

README.md and README.en.md are one document in two languages — line 3 of each
declares the other as its translation. They had drifted on the sentence that says
where the API list lives.

cabe4c6 (#203, docs(architecture): correct the ops method, the transaction types
and the API list scope) changed only the Chinese half: README.md now reads

API 一览见 docs/architecture.md，权威清单以 src/routes/mod.rs 的 router() 为准。

while README.en.md kept its pre-#203 wording (Full API reference: [docs/architecture.md](docs/architecture.md).). So the Chinese reader is sent to the
code and the English reader to a document that #203 was downgrading to an overview —
one question, two answers, from one pair of files that claim to be the same document.
git show --name-only cabe4c6 lists no README.en.md.

This PR fixes the value and gives the pair an enforcer.

Related Issue

None — found while scanning the repository's own documentation for cross-carrier
drift; the divergence is traceable to cabe4c6 and is not filed as an issue.

Changes

• Fix the value: README.en.md now says the same thing as the Chinese half — an
  overview link plus the authoritative source in code (router() in
  src/routes/mod.rs).
• Add src/readme_gate.rs, a test-only gate (#[cfg(test)] mod readme_gate in
  main.rs, so nothing enters the release binary): the two carriers must hold the
  same facts.
  - Compared positions are the language-independent ones only: inline code
  spans, markdown link targets, bare URLs, numeric literals. Each family is
  compared as a set; the expected value is derived from the files themselves
  on both sides — no snapshot, no hard-coded list.
  - Deliberately not compared (they legitimately differ between languages):
  prose, headings, table headers, emphasis, line order, line count.
  - Positive control: every family must be non-empty on both sides and the four
  together must yield enough items, so an extractor that returns empty sets
  cannot pass vacuously. Teeth: four synthetic arms, one fact dropped per
  family, each reported.
  - Scope, stated in the file: the gate is lexical. It proves the two sides
  carry the same structural positions; it does not prove the translation is
  faithful, and a competing fix that deletes the Chinese sentence would pass it —
  that judgement belongs to review, not to a gate.
• No configuration / data-structure changes, so config/config.example.toml needs
  no update.

Tests

• cargo test — 417 passed / 0 failed (baseline 413 on e567fae; +4 new tests)
• cargo fmt --check — rc 0
• cargo clippy --all-targets -- -D warnings — rc 0
• New unit tests added (src/readme_gate.rs)

A/B on the gate, both legs run locally on this branch:

tree	readme_gate result
README.en.md as of HEAD~1 (pre-fix)	the_two_readmes_carry_the_same_facts FAILED — exactly 2 divergences, both named: `router()` and `src/routes/mod.rs` 仅中文版有; the other 3 gate tests stay green
this branch	4 passed / 0 failed

Checklist

• Branch naming follows the convention (docs/…)
• Commit message uses Conventional Commits (docs(readme): …)
• Single responsibility, minimal diff (3 files, +299/−1; the one-line value fix
  plus one new test-only module and its registration)

[argszero]

Self-review (author is also the committer on this repo; allow_self_merge is on, so
GitHub will not accept an approving review from me — recording the review here instead).

Checked, in order:

1. The pair really is the only one. git ls-files | grep -iE '\.en\.|\.zh\.' returns
   exactly README.en.md, and README.md / README.en.md are both 80 lines and link each
   other on line 3. No other document in the repo has a language twin, so the gate's corpus is
   the whole family — not a sample of it.

2. The divergence is drift, not a choice. git log -1 -S'权威清单以' -- README.md → cabe4c6
   (docs(architecture): correct the ops method, the transaction types and the API list scope #203); git show --name-only cabe4c6 lists README.md and docs/architecture.md and no
   README.en.md. Pre-docs(architecture): correct the ops method, the transaction types and the API list scope #203 the two halves were aligned ("API 全清单见 …" ↔ "Full API
   reference: …"), so docs(architecture): correct the ops method, the transaction types and the API list scope #203 changed one side of a pair. The English file has never carried the
   authoritative-list sentence.

3. The fix says the same thing, not the same words. The new English clause keeps the
   overview link and names the authoritative source in code (`router()` in
   `src/routes/mod.rs`) — the two positions the Chinese half adds. It is not a literal
   translation and does not need to be.

4. The gate's expected value is derived. No list of tokens is written down: both sides are
   extracted from the files, and equality between the two extractions is the assertion. Only
   the language-independent families are compared (inline code, link targets, bare URLs,
   numeric literals); prose, headings, emphasis and line order are excluded by construction.
   A reviewer can therefore change either README freely as long as it stays the same document.

5. It can't pass vacuously. Four families must each be non-empty on both sides, and the
   four together must yield at least 30 items per file (today 64 across the pair) — an
   extractor that returns empty sets is caught. The four synthetic arms drop one fact per
   family and each is reported.

6. A/B on the pre-fix tree. Reverting only README.en.md to HEAD~1 makes exactly one
   test fail — the_two_readmes_carry_the_same_facts — naming exactly the two spans
   (`router()`, `src/routes/mod.rs`); the other three gate tests stay green. So the
   gate discriminates on this axis and not on the new code's mere presence.

7. CI: test / fmt / clippy pass (1m33s), msrv pass (29s) — the msrv job only builds, and
   the new module is #[cfg(test)], so it adds nothing to the release binary or to the
   declared-1.86 build.

Scope, restated so it is not over-read: the gate is lexical. It proves the two carriers
carry the same structural positions; it does not prove the translation is faithful, and a
competing fix that deletes the Chinese sentence would pass it. That is a reviewer's call, not
a gate's — and it is written into src/readme_gate.rs itself.