A case number names one case

Rationale

Three things assume the identity names the case: bash test/run.sh 96_worktree_topology runs one by name, test/cases/96_* is how a gate or a script selects one, and "case 96" is how every rule, claim, ADR and commit message in this repository refers to one. When two files carry an identity, the first is ambiguous, the second selects more than was meant, and the third points at two things and the reader cannot tell which.

Measured on 2026-09-12, across 812 refs: 265 distinct case files had ever been added under test/cases/, carrying 172 identities; 55 of those identities had been carried by more than one distinct file. The tree of origin/master held 206 files under 164 identities, and 25 of those identities were held by two to seven files at once — 97 by seven. Three sessions took 278 within minutes of each other that day, and a session that told its peers "284 is next" was wrong twice over: 283 was on a rescue ref and 290 on an open branch.

That is not carelessness. A session allocates from what it can see, and what it can see is the tree it branched from — the one place the next free number is least likely to be. The decisions had the same defect and majordomus adr next answered it: a survey of the working tree, every sibling worktree on disk, every ref, and the peer board, each saying what it claimed. Case identities are the second identifier allocated by hand here, and this rule gives them the same survey and a gate for the half a gate can see.

Required behaviour

Allocating. A new case takes the identity scripts/case-numbers next reports: one above the highest identity any surveyed source claims. The allocation is monotonic and never recycles a hole, because a hole is a number taken and withdrawn or taken and not yet written, and no survey can tell those apart. The identity is announced — the number, not only the path — before the file is written (project.work-is-claimed-before-it-is-built).

The survey is a best reading and says so. A ref scan sees only what was pushed; the board sees an announcement only while its connection lives, and is unreachable more often than not. A number two sessions take in the same minute, neither pushing nor announcing, is invisible to both, and no mechanism removes that: the two gaps are complementary. So next names every source it reached and every source it did not, lists the identities allocated and not landed on the base with who holds each, and never presents its number as a reservation (project.a-verdict-states-its-subject).

Holding. No two files under test/cases/ share the text before their first underscore. That text is taken verbatim, so 87 and 87b are two identities and not a collision: a case and its deliberate second half is a convention this repository already uses.

The collisions of the day the gate was written are recorded in .ai/repo/ci/case-numbers-baseline.txt and may remain, because they are spread across other sessions' branches and rescue refs, and renumbering them would break every one of those. Nothing may be added to it. A file joining a collision the baseline already carries is refused like any other, or the ratchet would slow the rot rather than stop it. Shrinking the list needs no permission: check names each entry that no longer collides. Rewriting the baseline to turn a refusal green is the one thing it is not for.

Failure behaviour

scripts/case-numbers check, registered as the case-numbers gate, exits 10 naming each new file, the identity it shares and what it shares it with, states the population it examined, and names scripts/case-numbers next. --strict refuses every collision, baseline or not, for whoever is paying the debt down.

scripts/case-numbers next exits 0 whenever it answered, including when the board or the refs could not be read — each is then reported as not surveyed, never silently omitted. scripts/case-numbers collisions makes the cross-ref state visible: every identity more than one file has ever carried, whether each file is on the base, and who carries it now.

Verification

test/cases/560_a_case_number_names_one_case.sh builds the repositories it measures:

• the gate over trees that are distinct, that collide, whose collision is on the baseline, whose baselined collision a third file joins, whose collision was renumbered away, and 87 beside 87b;
• the allocator over a repository whose highest identity is only in a sibling worktree's uncommitted file, whose next highest was added and deleted on a rescue ref and is carried by nothing, whose unmerged branch holds an identity the base lacks, whose base renamed a case that older branches still carry, and whose branch carries a different file under an identity the base holds — each of which a survey of fewer sources answers wrongly;
• a board document announcing an identity above every file, and a base that does not resolve, each of which must be said rather than absorbed.