Adds the library world's guards as corpus rows

[johnnyt]

What this lands

conformance/cases/library.json, the library example world's renewal and
standing guards as authored corpus rows, plus the regeneration they force.

The rows, enumerated, and what each pins:

1. library/renew-allowed - the renewal guard in full, satisfied.
2. library/renew-refused-by-hold - the same guard, a hold waiting.
3. library/renew-refused-at-max - the same guard, at the renewal limit.
4. library/renew-refused-blocked-standing - the same guard, standing blocked.
5. library/standing-good-no-fines - good standing and nothing owed.
6. library/fine-owed - the same guard with one unpaid fine.
7. library/due-soon-window - the due-soon reminder three days before the
   end of the loan period.
8. library/pickup-window - the seven-day pickup window as a duration.
9. library/overdue-days - an overdue copy not yet at the lost-item point.
10. library/max-renewals-boundary - the limit is exclusive at the boundary.
11. library/nested-patron - two chained property reads over a nested context.
12. library/holds-pending-negation - the holds flag negated.

That enumeration is twelve rows, and twelve library/* ids appear in the
regenerated tier files.

Every context is drawn from the example world: the patron ids, the copy and
branch ids, the renewal limit of 2, and the loan, pickup and lost-item
durations. No case carries a date.

Corpus hash

• before: sha256:548f54cacdcb700df0c47d67f86a944b5dbe6b0c6f96ef0c4c6fbb95b0494892
• after: sha256:bb60ec82faba2e82ef1e7abaac8d26257db2a77cd4ce0c083b683c5bc063fd43

mix corpus.generate ran the authored rows through the real compiler and
evaluator, which is what makes their expected values true rather than
asserted, and rewrote the tier 1, tier 3 and tier 4 files and the manifest.
The asserted tier on every row agreed with the computed one.

Two rows depart from their sketch, because the ISA refuses the shape

• library/due-soon-window was sketched as 21d - 3d. docs/isa.md
  section 5's subtract bullet gives clauses for number-number, Date-Date,
  DateTime-DateTime and Date/DateTime-duration only, so a
  duration-duration subtraction is a TypeMismatchError. The row computes the
  window over a duration context value's days field instead, and its notes
  field says so.
• library/overdue-days compares the days fields of two duration context
  values rather than the durations themselves. The same section leaves ordering
  between two maps unspecified for siblings, and a duration is a map at the
  value level. Its notes field says so.

Both were found by reading the ISA after the generator refused the sketched
shape, not by weakening a row to pass.

Provenance

One file beyond the authored case and the regenerated corpus changed:
conformance/examples/registry.example.json. Adding two tier-1 rows moves the
manifest's corpus_hash, which that example pins, and the example claims the
evaluator surface complete at tier 1, so it also gains the two new tier-1
evaluator entries. Predicator.Conformance.RatchetRegistryTest's pin test and
its R5 completeness test both name that file in their failure text; the
canonical encoding was round-tripped byte-for-byte before the edit and the
entries were re-sorted by the order rule 2 specifies. A corpus regeneration has
refreshed that same pin before, in bdbfdab.

conformance/README.md gains one sentence under
## How to add a case, without any Elixir, saying the library/* rows are the
family's shared example world and where its definition lives.

No changelog fragment

changelog.d/README.md says not to write a fragment for "test harness, corpus
tooling, or conformance fixtures". This is exactly that.

Gate

Full mix quality, quoted whole:

Running quality checks...

✓ Format: No changes needed (273ms)
✓ Compile: dev + test compiled (warnings as errors) (308ms)

Running analysis stages in parallel...

○ Doctor: skipped (:doctor not installed)
○ Gettext: skipped (:gettext not installed)
○ Sobelow: skipped (:sobelow not installed)
✓ Dependencies: No unused dependencies (347ms)
✓ Credo: No issues (3.1s)
✓ Tests: 2,914 of 2,914 passed, 95.5% coverage (3.8s)
✓ Dialyzer: No warnings (4.3s)

✓ All quality checks passed!

The branch's base is main's tip, and the tree committed is byte-identical to
the tree that run went green on, so the commit cites this run rather than
re-running it.

Refs: px-5l8

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.