[oddharsh]

#1635 and #1647 fixed what a hand crawl found, a page at a time. This keeps the crawl and fixes everything it flags today.

What changed

bun run browser:check (tools/check-browser.ts) opens all 76 manifest pages in Chrome and reads four things on each:

reading	what fails it
scroll	a scroller wider than its box at 320, 360 or 390px (gotcha 97)
shift	cumulative layout shift over 0.01 at 1280 or 390px, under 4x CPU throttle (gotcha 98, new)
intl	any Intl call before the reader acts (gotcha 92)
error	an uncaught exception during load

It replaces phone:check, which was the first of the four, and takes over its non-blocking CI job (now browser check). config/browser-budget.json holds four exceptions, each with its reason. An entry a page no longer needs is itself a failure, so the file can only shrink. --base <url> reads a deployment in place of the local build.

What --base https://aadhar.sh found that a local run can't. /around and /ledger scroll sideways in production at every phone width (103px and 3px out at 390px). Both passed locally, where the harness's empty KV leaves their tables empty. Each table scrolls in its own box now.

Everything else it flagged, fixed here:

	pages	fix
scroll at 320px, 3 to 37px	/restore, /lens, /garage/blueprint, compression, dyno, masonry, octane	min-width: 0 on a range input, tighter padding or wrapping under 340px
shift 0.086 (0.192 at 390px in production)	/garage/round-trip	the frame holds the window's height while the Markdown loads; the demo's output box holds its final height
shift 0.021, in 4 of 9 loads	/access	the empty header strip above the graph reserves its aspect ratio
Intl at load	/restore	names months and weekdays itself

Why

Each defect class was invisible on a fast machine with a wide window, and each came back to "nobody opened the page the way a visitor does". The two production-only findings are the sharper version: a check with the wrong data passes.

Understanding check

• Intended model: a visitor-facing defect that a browser can measure fails the PR that introduces it. A reading is trusted only if its planted defect was caught in the same run.
• Prediction: revert any one of this PR's page fixes and browser:check names that page and reading. Add a budget entry for a clean page and it reports the entry as stale.
• Likely misconception or failure mode:
  • A green local run means production is clean. It doesn't for data-backed pages, which is why --base exists and why three budget entries say "local fallback only".
  • Layout shift is deterministic. With three pages loading at once a paint can land early: /garage/pretext read 0.056 in one sweep and 0.005 in six loads alone. A reading over the limit is taken a second time and judged on the lower one; a script that moves content after paint does it on every load, so it still fails.
  • A machine-dependent shift can be budgeted. It can't. I gave /access a 0.03 budget because it read 0.021 here; the first CI run read 0.000 there and called the budget stale. The page is fixed instead (second commit), and a budget now counts as stale only under half the limit, so a page on the line can't flap.
• Evidence:
  • Controls, every run: a 600px block (scroll, at each width), a 300px block pushed in above the content (shift), a formatted number and a built formatter (intl), a thrown error (error). A miss exits 2.
  • With the page fixes stashed and a stale budget planted: 10 failures and 2 stale entries. With them: 0 and 0.
  • On the GitHub runner (first CI run): every control caught, 0 failures across 76 pages, sweep in 2m25s. Its one complaint was the stale /access budget above.
  • The table wrappers, applied to the live /around and /ledger DOM at 360px: main overflow 133px to 0 and 33px to 0, with the box scrolling.
  • /restore's hand-built dates match toLocaleDateString("en-US") for all 131 restore points. /access reads 0.003 at most in 12 loads.
• Remaining uncertainty:
  • /restore showed dates in the visitor's locale and shows en-US now.
  • The budget carries /serendipity's Intl use (event times in the visitor's zone), which is real and stays. The other three entries are local-fallback shifts that production doesn't have.
  • Production still shows pre-perf(pages): four pages stop shifting after paint, and nine stop paying for Intl and trig at load #1647 behavior on some pages until that deploy lands, so I couldn't take a clean production reading of the whole set.

Validation

• bun run browser:check: 76 pages, 0 failures, 0 stale budgets, every control caught.
• bun run lint, format:check, typecheck, derive:check, pages:check, check-wrangler: pass.
• bun run routes:check: 380 routes, 0 hard failures.
• bun run test: 1777 pass, 0 fail (Node 26). cal tests pass.
• Touches ci.yml: the phone-layout job becomes browser-check, still non-required with continue-on-error. Three contract tests that name the job and the node script follow.
• No /a/ URL moves.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UnguFuTGB1AyfrTeixtbCT

[github-actions]

Wire-size diff vs merge base

7fdfd6cfb38ab63b3f82ab9a30c0c4a516b58b34 → 9ca5306685ca89ae71d0cb31ff0c35df6ce52802

Brotli q11, the quality the edge serves. Every number here is deterministic:
an unchanged file produces no row. Advisory — this check fails on nothing.

Client assets

No change. 40 files, 245.02 KiB Brotli total.

Pages

Total Brotli: 718.80 KiB → 718.95 KiB (+0.15 KiB, +0.0%) across 77 files, 11 changed.

No single file moved by 0.13 KiB or more.

_{Plus 11 files under the 0.13 KiB floor, net +0.15 KiB — usually content-hash churn in /a/ references, counted in the total above.}

Worker bundle

299.34 KiB → 299.50 KiB gzip (+0.16 KiB, +0.1%)

Largest module deltas (raw bytes in bundle):

file	base	head	Δ
.build/src/worker/around.ts	15.04 KiB	15.10 KiB	+0.06 KiB (+0.4%)
.build/src/worker/ledger.ts	13.43 KiB	13.49 KiB	+0.05 KiB (+0.4%)
.build/src/worker/lens.ts	134.08 KiB	134.13 KiB	+0.05 KiB (+0.0%)
.build/src/worker/dyno.ts	8.84 KiB	8.89 KiB	+0.05 KiB (+0.5%)
.build/src/worker/updates.ts	12.64 KiB	12.68 KiB	+0.04 KiB (+0.3%)

Dictionary transport

Family dictionary acquisition: no change, 14.50 KiB q11.

pd/: 446 → 457 deltas, 525.78 KiB → 529.44 KiB (+3.66 KiB).

Head tags vs merge base

No change. 77 documents, every title, description, canonical, robots, og:*, twitter:*, alternate, JSON-LD type and oracle status identical.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	aadhar-sh	1ae47ff	Commit Preview URL Branch Preview URL	Nov 10 2026, 08:02 PM