fix(pm): H19's summary clause ends its own sentence

[zhuangjianguo]

Fixes #13650

summaryLine's H19 clause carried its terminating . inside the cross-repo branch
alone, so the ONE rendering that ended the sentence was the one where an unrelated
optional probe happened to fire. On an ordinary healthy pass — no shortfall AND no probe
— the clause stopped at card(s) and a lone space joined it straight onto H20's lead-in,
so two disclosures read as one sentence and the boundary was findable only by a reader who
already knew the row inventory.

What changed

One file: scripts/pm/check-half-states.mjs.

• H19's clause now terminates unconditionally, with the same . every other clause
  on the line uses.
• The two optional branches (shortfall, cross-repo) contribute only their own text. The
  cross-repo branch already ended in a period of its own, so its leading space and the
  join's trailing space move into the terminator rather than adding a second one: the
  probe shape renders resolves them. , never a doubled .. .

Everything else on the line is byte-identical

Rendered at the base commit and at head over the same six shapes, then diffed character by
character. Nothing is removed on any shape, and the single inserted character sits in
H19's terminator region on all six:

shape	base → head	delta
healthy (no shortfall, no probe)	3247 → 3248	one inserted .
shortfall, no probe	3583 → 3584	one inserted .
cross-repo probe, no shortfall	3498 → 3499	one inserted .
shortfall + probe	3834 → 3835	one inserted .
summaryLine({}, 0)	3283 → 3284	one inserted .
every optional sub-clause firing	5287 → 5288	one inserted .

The clause-boundary extractor is unmoved: the partition control (every window's clause
length summing to the whole line's length — no gap, no overlap) holds on all six shapes at
head, so adding the terminator inside H19's clause moved no boundary.

Live read-only fire, at this PR's head commit

PM_SWEEP_REPO=objectstack-ai/objectstack node scripts/pm/check-half-states.mjs at
59d5c22bb, exit 0, read-only. The H19 to H20 boundary now reads:

Blocker liveness (H19): targets resolved on 27 of 27 distinct `Blocked-by:` target(s)
named by open `pm:blocked` card(s). Dispatch liveness (H20 + H27): remote branch read on
8 of 8 distinct claimed branch(es) …

A no-shortfall, no-probe pass — the exact shape that used to run on. The old adjacency
card(s) Dispatch liveness occurs 0 times in that run's output.

Pins

Scoped to H19 via the clause extractor, so no neighbour's text can satisfy them. Eight
cases: the clause ends . on the healthy, shortfall, probe and both-branches shapes;
the probe shape carries no doubled period and closes exactly once; and two deliberately
whole-line cases on the boundary itself, because the adjacency between two clauses has no
single window to scope it to.

All four shapes are pinned rather than one, because the defect WAS that the shapes
disagreed — a pin written on the probe shape alone would have been green for the whole
life of the bug. The neighbouring whole-line endsWith('not a gate verdict.') cases could
not see this at all: they pin where the SENTENCE ends, which is H37's business, not where
this clause ends.

Ablations

Both legs mutated on disk with the anchor hit-count asserted (a zero-hit edit refuses
loudly instead of measuring an unmutated tree), and restored against HEAD with the
restored blob hash compared to the HEAD blob and git diff HEAD proven empty.

• Reverting only the terminator to its previous shape: exactly 4 of 1791 cases red —
  the healthy and shortfall terminator cases and both boundary cases. The two probe-shape
  cases stayed green, which is the measurement that shows why one shape was not enough.
• Deleting H19's clause outright: 22 red, including
  summaryClause: h19Blockers opens its own clause on every run. H19's presence control
  is still red under its own ablation and the other 16 windows' presence controls stayed
  green, so the per-window contract is intact.

Verification

Self-test, quoting its own verdict lines: ✓ check-half-states self-test: 1783 cases pass.
at the base commit, ✓ check-half-states self-test: 1791 cases pass. at head (+8).

Gate family derived from the actual diff with
node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack, run at
59d5c22bb — 17 path-derived families plus the 2 convention-triggered ones this change
kind incurs (it edits a gate script). All green, each read from the gate's own verdict
line with the exit code captured before any pipe, except two that decline to measure
locally and say so themselves:

• node scripts/check-test-completeness.mjs — PREREQUISITE NOT MET, exit 3: it grades a
  saved turbo run test log and there is none locally. NOT MEASURED, not a red; its own
  text says so.
• node scripts/check-partof-closing-keyword.mjs with no PR context — NOT WIRED,
  exit 2. Its self-test spelling (pnpm check:partof-closing-keyword, 28 cases) is green,
  and the gate was additionally run against this body with PR_BODY set before the PR was
  opened.

No changeset: nothing here is published from any package — the diff is scripts/pm/**
only, so the PR carries skip-changeset.

Notes for the reviewer

Card relationships, stated once here and carried by no commit trailer (this branch
squashes): the scoped-assertion scaffolding these pins are built on came from #13654, and
its adjudicated default was that the rendered summary line must not change — this card is
the one sanctioned rendered-line change deferred out of it. Nothing from #13654 or #13628
is undone here. The shortfall wording and the cross-repo branch, both from #11218, are
untouched apart from the whitespace that moved into the terminator.

Generated by Claude Code

---

Generated by Claude Code