fix(e2e): two anchors #2057 broke that 49 green checks did not see

[rubenvdlinde]

Follow-up to #2116, which merged while this was being measured. Two things it got wrong.

#2057 broke two anchors and every check passed

#2057 renumbered the kanban delta's scenarios from DASH-V1-006d/e to 006f/g, to clear a collision
with two scenarios the dashboard spec already had. It did not update the two citations in
tests/e2e/spec-coverage/kanban-board-keyboard-status-transition.spec.ts, which still named 006d and
006e.

The tests kept running. They kept reporting green. They cited scenarios that no longer existed. 49
checks passed, twice, across #2057 and #2066.

Nothing caught it because gate 46 does not read @e2e. Under @spec it would have failed the same
afternoon. This is the clearest example I have of the gap now filed as ConductionNL/.github#726, and it
is self-inflicted, which is why it is worth writing down rather than quietly fixing.

Both are repointed at the canonical openspec/specs/dashboard/spec.md#scenario-dash-v1-006f-… and
…006g-…, per the convention #2116 landed. Each fragment was probed through the real helper before
it was written.

The measurement in #2116 was wrong, and wrong the way everything else here was wrong

#2116 claims "6,244 anchors, zero unresolvable". That was measured with a reimplementation of gate
46's resolver rather than the resolver, and it was wrong in both directions.

Mine missed two things the real one does:

• the flat openspec/specs/<cap>.md spelling, which predates the directory form and which the
  helper explicitly normalises. That alone made mine call 13 of planix's healthy anchors dangling.
• fragment checking. The helper verifies that a #fragment names a heading somebody actually
  wrote. Mine stopped at the file.

Run properly, dossiq has 28 unresolved @e2e anchors, every one of them anchor not found rather
than a missing file. That is exactly the opencatalogi#85 shape gate 46 was built for, surviving in the
half of the corpus the gate does not read.

	before	after
unresolved @e2e	28	26
unresolved @spec (real gate, real scope)	0	0

The 26 that remain are pre-existing and stay. Each needs a judgement about which scenario it meant,
which is the work rather than the sed, and the gate fix filed upstream surfaces them on the next PR
touching those files. Tracked as task 5.3 on the still-open spec-anchors-name-the-capability change,
which is why that change is not archived.

Upstream, filed with the fleet measurement

Taken with the real helper across all 21 fleet apps, by rewriting @e2e tags to @spec in a probe
file outside each repo and running check_spec_anchors.py with cwd set to the app:

• gate-46 reads only @spec: 1,964 @e2e anchors fleet-wide are checked by nothing, 194 of them dangling .github#726 — gate 46's pattern is @spec\s+(openspec/...), so @e2e is never
  read. 1,964 such anchors fleet-wide, 194 dangling, concentrated in 6 repos. Against 55,720
  @spec anchors with 1 dangling: the gate works, it just does not look at a fifth of the corpus.
  Recommended warning-only first, with a deadline rather than an open one.
• gate-46 never enumerates appinfo/ or scripts/: 52 anchors unopened fleet-wide, 8 of openregister's 9 in routes.php are dangling .github#727 — appinfo/ and scripts/ are never enumerated. 52 anchors
  unopened, and 8 of openregister's 9 in appinfo/routes.php are dangling, two of them ending in
  a stray full stop that was never a path.

The rule underneath all of it

Stop reimplementing the instrument. Every wrong number in this investigation came from a hand-rolled
resolver: 420, then 2,955, then "zero unresolvable". Every right one came from running
check_spec_anchors.py. It is the same failure three times, and the third time it was mine after I had
already written up the first two.

[github-actions]

Quality Report — ConductionNL/dossiq @ 7b81b63

Check	PHP	Vue	Security	License	Tests
lint	✅
phpcs	✅
phpmd	✅
psalm	✅
phpstan	✅
phpmetrics	✅
eslint		✅
stylelint		✅
build		✅
check-manifest		✅
check-vue3-compile		✅
test-l10n		✅
format		✅
check-schema-l10n		✅
check-l10n-js		✅
composer			✅	✅ 106/106
npm			✅	✅ 549/549
app:check-code	⏭️
info.xml	✅
REUSE	❌
PHPUnit					✅
Newman					⏭️
Playwright					❌
Hydra gates					✅

---

Quality workflow — 2026-09-09 11:36 UTC

Download the full PDF report from the workflow artifacts.

[rubenvdlinde]

CI

302 passed, 4 failed, 1 flaky. The 4 are identical to development's own set at fe1fa34f
(run), the same four that have been
red all day:

case-actions-menu:332   case-type-authoring-extras:494
contacts-domain:362     case-type-authoring-extras:645

The flaky one is case-type-authoring-extras:326, failed once and passed on retry, excluded by keeping
only specs with no passing attempt so it reconciles with the runner's 4 failed / 1 flaky rather than
reporting five.

47 checks, not 4: this branch is MERGEABLE, unlike #2116 before its merge.

The two tests whose anchors this fixes both ran and passed

✓ a case card exposes a keyboard-operable "Move to…" menu   (12.x s)
✓ case cards remain draggable for mouse/touch users         (10.3s)

Which is the whole point. They were passing before too, while citing scenarios that had not existed
since #2057. A green test proves the behaviour; it says nothing about whether the requirement it claims
to prove is still there. That is the job of gate 46, and gate 46 does not read @e2e
(ConductionNL/.github#726).

Anchor state after this PR

	count
unresolved @spec, real gate over its real scope	0
unresolved @e2e, real resolver over all 284	26 (was 28)

The 26 are pre-existing, tracked as task 5.3 on the open spec-anchors-name-the-capability change.
Each needs a judgement about which scenario it meant, so they are deliberately not swept here.