Open the 2.3 cycle, and record the ledger roster AGENTS.md missed

[derek73]

Release checklist steps 8 and 9, run after 2.2.0 went to PyPI.

Opening the cycle

• tools/differential/expected_since_2.2.0.toml, opened empty. DEFAULT_BASELINE and the ledger are coupled — _allowlist_for treats a missing file as a hard error, on the reasoning that an absent ledger classifies nothing and would make every diff report as unexplained — so the baseline cannot advance without it.
• DEFAULT_BASELINE 2.1.0 → 2.2.0 in compare.py.
• VERSION = (2, 3, 0), PRE_RELEASE = "dev", so an install from master normalizes to 2.3.0.dev0 — above 2.2.0, below 2.3.0, and never able to masquerade as the release it precedes. The docstring's own worked example moved with it.
• * 2.3.0 - Unreleased opened at the top of the release log.

CITATION.cff deliberately stays at 2.2.0: it names the last citable released artifact and is meant to lag for the whole cycle.

The roster AGENTS.md missed

Step 8 named four rosters to enrol a new ledger in. I hit a fifth — _CROSS_RULE_WINNERS — as a red run after the other registrations were already in, which is the worst time to find it.

Rather than just add the missing line, I measured which rosters are actually mandatory:

roster	keyed by	required for every ledger?
_SPAN_BEARING_RULES	filename	yes (equality)
_CORPUS_CLAIMS	filename	yes (equality)
_CROSS_RULE_WINNERS	filename	yes (equality, since #452)
_HONORIFIC_SOURCES	rule content	only if such a rule exists
_LATIN_ALTERNATION_SOURCES	rule content	only if such a rule exists

Step 8 said "enrol in _CORPUS_CLAIMS always, and in the others where they apply", which is wrong for two of the five, and it had _HONORIFIC_SOURCES filed among the mandatory ones. Rewritten under Required, by filename / Conditional, by rule content headings, with _CROSS_RULE_WINNERS added and #452's reason for its equality check recorded — a ledger with no rows had been indistinguishable from one needing none.

Each roster does fail loudly and name itself, so this is self-correcting; but it is three separate red runs if you add them one at a time, which is what the rewrite is meant to prevent.

Also records, at _CORPUS_CLAIMS, what a pure corpus rename looks like there — names unchanged, roles unchanged, digest moved. That came up renaming de Mesnil Juan → Jean during the release, and the guard's own message explains a growing count but says nothing about a moving digest.

Verification

• 6,198 tests; mypy and ruff clean; docs build clean
• Gate against the newly published 2.2.0: intentional diffs: 0, unexplained: 0 — master is byte-identical to what shipped, the correct reading on day one of a cycle
• 1.4.0 and 2.1.0 still exit 0, at 226 and 106 intentional diffs

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.66%. Comparing base (e52aeaa) to head (b3e1e77).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #480   +/-   ##
=======================================
  Coverage   98.66%   98.66%           
=======================================
  Files          45       45           
  Lines        3218     3218           
=======================================
  Hits         3175     3175           
  Misses         43       43           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

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