An anchor that cannot be wrong is not an anchor

tri issues numbers --as-of 2027-01-01 printed 486 open issues — today's count — under the heading AS OF 2027-01-01T23:59:59Z. It looked anchored, read as history, and was a clock reading wearing next year's label.

GitHub does not object. created:<=2027-01-01 composed with closed:>2027-01-01 answers 486 + 0 = 486 without complaint, because every issue that exists was created before next year and none has been closed after it. The query is well-formed and its answer is worthless.

§461 refused a malformed date on the grounds that a date silently becoming "today" is worse than no anchor. A future date does exactly that while parsing perfectly, so the refusal has to be about the day being CLOSED, not the string being well-shaped. Today is refused for the same reason as tomorrow: its end is still in the future, so the count differs from itself by evening.

$ tri issues numbers --as-of 2027-01-01
Error: --as-of 2027-01-01 asks for a day that has not ended (today is 2026-09-03 UTC).
The answer would be TODAY'S count under a heading that reads as history, which is worse
than no anchor at all -- GitHub answers such a query without complaint.

The calendar is load-bearing on two days in a hundred thousand

The comparison needs today's UTC date, so a civil calendar is transcribed rather than invented, and tested on the dates that break a wrong one: the epoch, the day before it, a leap day, and the day after February in a century that is not a leap year. Swap 36_524 for 36_525 and it invents 1900-02-29, a date that does not exist. Exactly two inputs in a hundred thousand discriminate — 1900-03-01 and 2100-03-01 — and they were found by sweeping the mutant against an independent calendar, not by picking dates that looked interesting.

One expectation in that test was wrong when written: civil_from_days(20_699) guessed 2026-09-04, and the answer is 2026-09-03. The code was right and the test was mine. That is what an independently derived expectation is for.

--limit was one of two spellings

The previous PR closed four --limit call sites and called the class closed. The same class is written per_page= in a URL, and that is 22 more lines.

tri gates fetches now walks the crate and reports it:

  of those, FETCH SITES         24
  refused as not a fetch        25   (--excluded prints them)

  complete by construction:
    --paginate                  5
    per_page=1 + total_count    2

  bounded, and:
    asks whether the page filled  6
    per_page=1, NO total read     2
    prints what it got            9

The nine are gates unmeasured, gates dead, red now and four in prcheck. None bites today — 62 workflows against a page of 100, 35 check-runs against 100 — and the margin that matters is still the issue one, 486 against 500. §462 is corrected in place.

Grepping one spelling and calling the class closed is the same error as counting one population and calling it the subject.

The census counted itself, twice

Its first rule was "the line names per_page= and names repos/", and it matched its own definition, which names both. The fix is not an exception list: both needles must live in the SAME string literal, which a rule never has and a URL always does.

Then it counted its own test fixtures — 25 sites where a hand count said 24. Excluding test modules looked like one line, everything after the first #[cfg(test)], and that was checked rather than assumed: five files in this crate carry real top-level functions AFTER their test module, gates.rs fifteen of them. The module has to be closed by a } in the first column, and the attribute has to be seen before the mod — or main.rs's forty ordinary module declarations put the walker in test mode for the rest of the file, silently dropping every fetch after the first.

What made both visible was the LIST, not the total. A census that prints only its count cannot be checked; one that prints its members can. --excluded prints the other half of the reading: 25 lines name a spelling without fetching, against 24 that fetch.

• fourteen clauses mutated, fourteen killed
• cargo test -p tri — 515 passed, 0 failed
• skill §463, §464; §462 corrected in place

Refs #2994