The board is not scannable: machine-generated titles dominate it, and every row carries the same weight

[bdelanghe]

From the installed app. The board renders correctly and reads badly — the failure is in what it emphasises, not in what it fetches.

What is wrong, concretely

1. The link underline is hand-rolled on a wrapping block.

.row__title a { color:inherit; text-decoration:none; border-bottom:1px solid var(--line); }

border-bottom on an inline element that wraps to four lines draws four disconnected segments. It reads as broken rendering. text-decoration with text-underline-offset follows the text properly; better still, make the row the target and drop the per-title rule.

2. Machine-generated titles are shown raw.

chore(deps): bump bounded-systems/.github/.github/workflows/_claim-sweep.yml from 14c1a9b3ade8bb6416f46c890eda21823647da09 to d126d721fa50275e12af0bdeaf1a2ff3016eedfd

About 90 of those ~150 characters are hex no one can read, and the two full SHAs push to the viewport edge. A dependabot SHA bump is the least decision-worthy row on the board and is consuming the most space on it. The board exists to answer what should I pick up — noise that cannot be picked up should not outweigh work that can.

The transform is semantic, not cosmetic: bump <file> from <sha> to <sha> → the file, and 14c1a9b → d126d72. Same information, one line.

3. No hierarchy. Title, repo and number are near-identical in weight, so nothing is scannable — the eye has no entry point per row and the sections read as one undifferentiated column.

4. Overflow. Long unbroken hex reaches the edge and clips; nothing constrains it.

Direction

Adopt the semantic/accessibility-driven line the org already has rather than inventing a local style: @bounded-systems/brand tokens for colour and type, lone's per-component a11y spec for the interactive parts (the row target, the enable button, focus states). Tap targets are 44px minimum; the enable button already honours that and rows do not.

Explicitly not proposed: hiding dependabot PRs. The board vouches for a count ("the first 5 of 6"), and hiding rows would make that count a lie. Demote and compress, do not filter.

Related

Not a rendering bug — the data layer is fine. #12/#13 are about freshness, #22 about how the board is fed; this is about what the board does with what it already has.

[bdelanghe]

Measured against the live site by driving it in headless Chromium at a pinned 390×844 mobile viewport. This replaces the eyeballed claims in the description — one of which was wrong, and the correction is the more serious finding.

The board is being rendered zoomed out, and that is the root cause

The meta viewport is correct:

width=device-width, initial-scale=1, viewport-fit=cover

but the page reports innerWidth: 493 on a 390px viewport. Nothing "overflows" — the browser widened the layout viewport to fit the content instead, then scaled it down. Everything on the board therefore renders about 21% smaller than designed. That is why text appears to run to the very edge and look clipped.

Adding one rule live drops the layout viewport straight back to the real width:

.row__title, .row__title a { overflow-wrap:anywhere; word-break:break-word }

before      : 493
after wrap  : 390

The cause is the unbreakable 40-character hex in dependabot titles: it sets a minimum content width no ancestor can satisfy, so the viewport expands to accommodate it.

Correction to the description: it claimed the hex "reaches the edge and clips". It does not clip — nothing overflows, because the layout widened instead. The visible symptom is the same but the mechanism is different, and it matters: an overflow-x fix would have changed nothing, because there is no overflow to hide.

Space actually spent

	rows	avg height	total
chore(deps)	5	175px	876px
everything else	10	116px	1158px

A third of the rows are dependabot and they consume 43% of the board's vertical space, each one averaging 1.5× the height of a row someone could act on. The longest title measured 167 characters in a 195px-tall row.

Also observed

navigator.serviceWorker.getRegistration() returns null under an iOS user agent. The client returns early on the "add to Home Screen" branch before registering the worker, so iOS browser visitors never register it and never precache /offline. Harmless for push (iOS only grants it to installed apps anyway) but it does mean the offline page does not exist for those visitors. Separate from #60; noting it here rather than opening a third issue.

Reproducing

Headless Chromium needs --ssl-version-max=tls1.2 to get through this session's egress proxy — its TLS 1.3 ClientHello gets the tunnel closed mid-handshake. That caps only the browser→proxy leg, which is re-terminated anyway; certificates are still verified.

Caveat on all of the above: this is Chromium, not iOS Safari. The layout arithmetic and the measurements transfer; standalone-mode behaviour and apple-touch-icon capture do not, and still need a real device.

---

Generated by Claude Code

[bounded-systems-front-desk]

🎫 Claimed by claude/bounded-desk-tools-3fcown — dispatched by bdelanghe — run

Authorization: human-authorized · assurance aal2
Authorized by bdelanghe with credential Si98d1nbmy2LL3nSc1--ROWWL8E (relying party keeper.bounded.tools).
Claim-request digest (the passkey's challenge, recomputed by this door): 86f0e4600188eae83633b3ea8e29e2c8f3aebeafcf0eb135d5bb822608f2d6b8 · base64url hvDkYAGI6ug2M7PqjiniyPOuvq_PDrE11buCJgjy1rg

Release by removing the claimed label (and unassigning) if work stops without a merged PR.

[bdelanghe]

🔓 Released — label removed and unassigned. #62 merged the design pass.

Landed: the root cause (.row__title { overflow-wrap: anywhere } — measured 493 → 390 at a 390px viewport, and 320 at 320px), semantic title compression (167 chars → 35 on the worst live row), routine rows down from 43.1% of vertical space to 19.2%, <ol> of single block links with role="list" and accessible names carrying repo and number, sections as labelled landmarks, and the type/space scale taken from @bounded-systems/brand.

Deliberately still open:

• The liveness half — a visibilitychange / pageshow refresh for the resume gap. iOS restores a suspended app's DOM without re-fetching, so the board can be hours stale with neither the HTTP cache nor the service worker involved. Scoped as its own change and not in Take colour from the brand package, and admit the lanes that need to notify #62.
• The layout suite is not gated. standard.yml runs node --test test/*.test.mjs; test/layout/*.test.mjs runs in no workflow. Every assertion above — innerWidth, touch targets, forced-colors bands, list structure — is protected by nothing at merge time. Both adversarial reviewers found this independently. Worth its own issue and PR rather than bolting a browser into CI unreviewed.
• Two rendering nits reported and not fixed: the overview still passes a marker where renderPrs dropped it as duplicative, and SECTION_COPY sets markerLabel: "Status" for claims, contradicting the row doc comment about what that slot means. Both currently inert.

Also worth recording: the live site now reports 390, not the 493 in the description — because the feed changed, not because the defect went away. The board republished at 01:29Z and the longest unbreakable run is currently 21 characters. The defect is latent, which is exactly why compression is not a substitute for the wrap rule: the feed is not ours.

---

Generated by Claude Code