Triage needed: 41 bare em-dash fallbacks across 33 files are NOT one class — decide the taxonomy before any lint rule or bulk sweep is written

[os-justin]

Filed by the domain:ui PM seat (session_01YBWFb5YgMU5dw8p2VKj16S) from the objectui#8491 dev's widest census. ⛔ Not claimed. This card asks for a decision, not a fix.

What was measured

While landing PR #8503, the objectui#8491 dev ran four instruments of widening breadth over origin/main, each with a lit control:

instrument	result
1 — the card's own class string	2 files, 4 sites — reproduces the card exactly, which is the warning sign: it grepped for the string the card had already found
2 — any muted-foreground/50 placeholder	5 sites, incl. a fourth site in ObjectGrid.tsx itself the card missed
3 — class-agnostic bare em-dash span	13 sites / 11 files — premise refuted
4 — bare em-dash ternary/nullish fallback, no span	41 sites / 33 files — refuted decisively

PR #8503 took instrument 3's population from 13 to 8; objectui#8504 addresses six of the remaining eight. Instrument 4's 41 are untouched and are the subject of this card.

(Independently re-measured by this seat with a narrower single-line regex: 9 → 6 on instrument 3. The counts differ by regex breadth; the direction and the conclusion — the class shrinks and survives — hold on both.)

Why this is triage and not a sweep

The 41 are not one class. Sampled from the census, they include:

• a Select option label (an em-dash as a menu entry, not an empty value)
• a duration fallback (a formatting default, not an absent field)
• codegen writing markdown cells (an em-dash as table syntax in generated text)
• four inside ObjectGrid.tsx's own mobile card layout, which — measured by rendering, not read — omits empty fields before the cell renderer is called, so the placeholder there is on a path that never draws for an empty value

A bulk swap would put EmptyValue into a <select> and into generated markdown. A lint rule written now would need an exemption list longer than the fix.

The decision this card asks for

Partition the 41 into:

1. genuine empty-value placeholders → the objectui#8504 repair applies
2. decorative / syntactic em-dashes (option labels, markdown, separators) → out of scope permanently; they must be nameable so a future rule can exempt them by shape, not by path
3. unreachable placeholders (like the mobile-card-layout four) → a separate question about whether the dead branch should exist at all

Only after (2) has a stable definition is the lint-rule question answerable — "forbid a bare em-dash placeholder outside packages/components" is deliberately NOT filed as a card yet, because its scope is exactly what this triage produces. Filing it first would invert the order and produce a rule nobody can enforce (ADR-0049: enforce-or-remove).

Method note for whoever does the triage

⚠️ Instrument 1 is the cautionary tale and it generalises. Grepping for the class string a card already quotes reproduces the card's own census and confirms it — that is not measurement, it is an echo. Every widening instrument here found sites the narrower one could not see, including one in the very file the card was about.

⚠️ git grep -- 'packages/*/src' is a full-path glob, not a directory prefix: it silently matches nothing. Use 'packages/*/src/*', and note it excludes apps/ — census that separately.

⚠️ Truncated output is not a reading: head -4 on ImportWizard.tsx showed only its i18n label decorations and hid the real carrier at :1034.

Related

objectui#8491 / PR #8503 (where the census was produced) · objectui#8504 (the six accessible-name carriers) · the DetailSection / HeaderHighlight card (the aria-labelled pair) · objectui#8475

Dedup

⚠️ Declared rather than claimed. This card is a request to triage a population measured today; no prior card states that population. Single-token searches for AiPendingActionsInbox, GridSchema and HeaderHighlight (run for the sibling cards) surfaced nothing adjacent. Suggested query for a triager: EmptyValue.

[os-justin]

Correction to the instrument note in this card and in objectui#8504 / objectui#8505 / objectui#8506

Those four cards carry a note from me saying, in effect: "multi-token free-text queries on this repo's issue search return dead zeros; a single distinctive token works."

The second half is too strong and I have now measured a counterexample.

search_issues for ValueDataSource — one distinctive CamelCase token, repo-scoped, exactly the shape I recommended — returns total_count: 0, while objectui#8447's own title reads "ValueDataSource's $-dialect matcher waves through every operator it does not recognise…".

Index lag is not the explanation: InspectorSelectField returns 3 hits including objectui#8450, which was filed later than #8447.

So the accurate statement is stronger and less convenient:

A zero from this repo's issue search is not evidence of absence, whatever the query shape. A firing lit control proves the index is reachable, not that it is complete. A non-zero result is usable; a zero is not.

What this changes in practice:

• A dedup section may claim "searched, found these neighbours, they are distinct" — that claim rests on results that came back, and stays valid.
• A dedup section may not claim "searched, nothing matching, therefore no duplicate" on the strength of a zero. Declare it instead, the way objectui#8513 / objectui#8514 / objectui#8515 do, and name a manual check for the triager.

The four cards' own dedup conclusions are unaffected — each was backed by a search that returned results I read (AiPendingActionsInbox → 1, GridSchema → 13, HeaderHighlight → 2, KanbanImpl → 1). It is the generalised advice in the note that was wrong, and it is corrected here rather than left to mislead the next triager.

This belongs on this card specifically because this card is the methodology one: it asks someone to census a 41-site population and will be read by whoever writes the eventual lint rule.

---

Generated by Claude Code

[os-justin]

Measured addition to this card's method note: no single probe in this thread found every carrier

The objectui#8504 dev found a seventh carrier of the accessible-name class that three prior censuses missed, and the reason generalises past "widen your regex". I re-measured it on origin/main (tests excluded) before writing this.

The two probes have disjoint blind spots

probe	finds ObjectGrid.tsx:4362	finds AgentPreview.tsx:302
muted-foreground/50 (instrument 2)	yes — this is how the fourth site was found	no — that file uses text-muted-foreground, no opacity suffix
>—< (element-agnostic)	no — its content is {t('grid.empty')}, not a bare em-dash	yes
every <span-anchored probe (instruments 1, 3, and mine)	no	no — it is a <div>

packages/app-shell/src/views/metadata-admin/previews/AgentPreview.tsx:302
  if (rows.length === 0) return <div className="text-muted-foreground italic">—</div>;

git grep -n '>—<' -- 'packages/*/src/*' 'apps/*/src/*' returns 10 carriers: ObjectGrid's three (closed by PR #8503), objectui#8504's six, and this one. It does not return ObjectGrid.tsx:4362, which PR #8503 also fixed.

So the two instruments that between them found every carrier each returned a false zero on the other's unique hit.

What this changes for the triage this card asks for

The census that partitions the 41 must not be a single grep, however wide. Every instrument in this thread was anchored on something — a class string, an opacity suffix, an element name — and each anchor's blind spot was invisible from inside that instrument. A zero is byte-identical to absence in all of them.

Run at least three orthogonal probes and reconcile their differences, not their union:

1. element-agnostic content — >—<, and the {t('…')}-content variants it cannot see
2. class-family — text-muted-foreground with and without an opacity suffix
3. fallback-shaped — ?? '—' / : '—' with no wrapping element at all (instrument 4's 41)

Where two probes disagree about a file, that file is the finding.

Consequence for the lint rule this card gates

A rule written against one anchor inherits that anchor's blind spot and will read as enforced while missing a whole element shape — the precise failure ADR-0049 exists to prevent. Whatever shape the rule takes, it needs a test proving it fires on all three of <span>—</span>, <div>—</div>, and a bare ?? '—', or it is not the rule it claims to be.

Filed for the seventh carrier itself: objectui#8520 (a block-level empty-collection state, not an inline missing-value placeholder — a genuinely different scope question, which is why it is not a mechanical repeat).

---

Generated by Claude Code