fix(tooling): derive check-wildcard-fallthrough's self-test case count from its own ledger

[baozhoutao]

Fixes #15231

The self-test verdict of scripts/check-wildcard-fallthrough.mjs printed a
transcribed 17 cases while the body ran 18 assertions and the battery roster
pinned 18. Nothing derived that number and nothing compared it — the fourth
instance of a class already filed twice (#13963, #13536) and once more in
#16007. Per triage, the route is option 2: derive the count, not correct the
literal.

Before / after

$ node scripts/check-wildcard-fallthrough.mjs --self-test     # origin/main 0a61db1f5d
✓ self-test: 17 cases                                          exit 0

$ node scripts/check-wildcard-fallthrough.mjs --self-test     # this branch
✓ self-test: 18 cases                                          exit 0

$ pnpm check:wildcard-fallthrough                             # this branch
✓ self-test: 18 cases
✓ wildcard fall-through: 8 yielding / 0 ratcheted / 5 exempt (13 namespace-claiming mounts)
                                                               exit 0

Re-derived on this branch's base, not taken from the card: the literal was at
scripts/check-wildcard-fallthrough.mjs:634, and grep -c "^\s*assert(" over
that file returns 18.

Which number is printed, and why that one

The line now prints assertions that actually RAN this run — the sum of
batterySeen, the same ledger the battery floor immediately above already
evaluates. It does not print assert( call sites in the source.

Today both facts are 18, so a reader would otherwise assume whichever one was
not meant. The runtime ledger is the honest choice for a line whose words are
"cases": a call site that never executes is not a case that held, and the whole
reason this class exists is that a printed number must be produced by the run
it describes. It also makes the printed number and the floor structurally
incapable of disagreeing, since they read one ledger.

Summing the whole ledger is exact rather than approximate: reaching the verdict
means the set-difference check above found every battery that registered to be
a declared one, so the sum has no unattributed remainder.

No new floor is introduced. SELF_TEST_BATTERIES keeps its existing pinned
value of 18 and is untouched — this change does not convert a printed number
into a gate.

Two comments rewritten, deliberately

Both would have become false on landing, standing next to a derived line:

• the in-place comment above the literal explained that 17 was knowingly
  wrong and "left as-is here so this change stays a pure no-op on output" —
  that sentence was about the PR that added the floor, not this one;
• the roster's rationale block described the verdict, in the present tense, as
  spelling a transcribed count.

Both now describe what the file does.

Nothing outside the file pinned this stdout

git grep -n "17 cases" at the base returns 10 hits: CHANGELOG prose and test
fixtures about unrelated 17-case corpora, plus this file's own two lines. None
is an expectation of this file's output. The instrument was controlled — the
same pattern scoped to the file finds line 634 — so the zero is a real zero.
docs/audits/2026-09-self-test-shape-census.md names this script but records
its failure line, not the success line. This change deliberately moves output,
so it could not have ridden the byte-identical batch that left the literal wrong.

Verification

Positive and negative controls, each mutated on disk (blob hash shown to differ
from the HEAD blob), run, then restored under a trap and proven restored by
hash equality plus an empty git diff HEAD:

leg	mutation	printed
baseline	none	✓ self-test: 18 cases (exit 0)
positive control	one deliberately added 19th assertion	✓ self-test: 19 cases (exit 0)
negative control	an assertion's message text edited, case count untouched	✓ self-test: 18 cases (exit 0)
deletion	one assertion removed	floor reds: registered 17 case(s), below its pinned floor of 18 (exit 1)

And the same positive control against the pre-change code, to show what the
literal could not do: origin/main's version of the file plus a deliberately
added 19th case still prints ✓ self-test: 17 cases, exit 0.

Restore proof for every leg: git checkout HEAD -- path, then
git hash-object path equal to git rev-parse HEAD:path
(bf00750b2d11cb8b0af7c4c0216d9c355fca5e4a) and git diff HEAD empty.

Gates — node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack
derived 31 commands; all 31 were run and all exited 0 (exit code captured before
any pipe), reconciled with --ran:
✓ dispatch-gates --ran: 31 derived famil(ies) accounted for — 31 run, 0 NOT-MEASURED.

Repo-wide lint at final HEAD 3185352ffc: pnpm exec eslint . --no-inline-config --format json — 6309 files, 0 errors, 0 warnings, exit 0. No narrowing claimed.
Control-character self-scan over the changed file is clean, and
pnpm check:nul-bytes is green.

Changeset

skip-changeset applies and is set on this PR: the diff is one repo-internal
gate script under scripts/, which publishes from no package.

---

Generated by Claude Code