[oddharsh]

Stacked on #353, which found these. This fixes them.

1. The photo grid was twelve blank squares

Every tile returned HTTP 200 and none decoded, because the tiles are AVIF and Kitesurf cannot read it. A request-level view calls that healthy, which is why nothing here caught it.

lib/photo-grid.js names one URL per tile deliberately, and that decision is untouched. This adds a recovery path instead, and the design turns on a fact measured first: Kitesurf does fire error on the tile it cannot decode. One delegated listener parses the stem back out of the failing URL and retries through /images/<stem>.jpg, which 301s to the hashed JPEG. No bytes are added to any tile.

Verified against the live grid with the listener as shipped:

KITESURF before  12 tiles, 12 broken,  0 repaired
KITESURF after   12 tiles,  0 broken, 12 repaired
CHROME   after   12 tiles,  0 broken,  0 repaired

The last row is the point: a browser that decodes AVIF never fires the event, so the fast path costs nothing.

The <noscript> twin gets <picture> instead, and the asymmetry is deliberate. Inert markup cannot instantiate a fallback, so the churn that took <picture> off the live tiles is impossible there, while a no-JS reader has no script running to repair anything.

Worth knowing for future decisions: <picture> works correctly in Kitesurf. /photos still ships it, and Kitesurf selected the JPEG for all 158 tiles and broke none, because a well-behaved engine declines a source type it cannot decode rather than claiming it. Gotcha 7's decode-failure warning does not apply to this engine.

2. Four surfaces advertised a twin and served none

/writing, /lens, /coffee and /around all carry flags.agents: true and had no Markdown twin. All four render from the Worker, so htmlFileFor() returns null and there is nothing for the generator to read.

/writing is generated from www/writing/posts.json, the same committed registry the page renders. A hand twin would go stale on the next post; this is the ordinary twin argument reaching one step further back than the HTML. Each row links to the post's .txt, so an agent gets prose with no Notepad chrome.

The other three are hand twins beside bot.md and terminal.md, each stating plainly that it describes rather than mirrors, and pointing at the JSON that carries the changing part. Every fact was read off the live endpoints rather than recalled.

Three carry a warning an agent needs that the HTML only implies: booking is a request, so a successful POST means "asked" and never "booked"; availability fails closed, so a 503 is the safe answer; a crawl skip is a legitimate outcome.

All four are swept by the route oracle now (134 routes), each with a load-bearing sentence as its marker rather than a title.

3. WebMCP, investigated rather than fixed

Not actionable here, but it resolves the open unknown from #353.

• Kitesurf gets no bridge injection at all. No <script src="/.webmcp/bridge.js">, no request for it. The browser tool catalog is simply absent on the agent browser.
• Kitesurf ships a stub document.modelContext carrying _listeners and _ontoolchange and no methods.

That second one is the same shape as the navigation stub that broke nav.js, which makes it a bug class rather than a one-off: this engine ships stub objects for APIs it does not implement, so if (window.X) is unsafe by default and if (typeof X.method === "function") is the only reliable form.

I swept the client scripts for other instances. tooltip.js touches Temporal behind a bare typeof check but wraps both call sites in try/catch with a Date fallback, so a partial implementation degrades correctly. Everything else is either our own modules or already checks the method. navigation was the only unguarded one and #353 fixes it.

249/249 contract tests, typecheck clean, route oracle 134/134, page contracts clean.

[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	Updated (UTC)
✅ Deployment successful! View logs	aadhar-sh	e4d1de8	Aug 14 2026, 02:50 PM

[github-actions]

Wire-size diff vs merge base

d0dabd5 → aeaab7f

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. 19 files, 68.08 KiB Brotli total.

Pages

Total Brotli: 494.22 KiB → 494.35 KiB (+0.13 KiB, +0.0%) across 48 files, 1 changed.

file	base	head	Δ
index.html	7.78 KiB	7.90 KiB	+0.13 KiB (+1.6%)

Worker bundle

278.56 KiB → 278.57 KiB gzip (+0.01 KiB, +0.0%)

Largest module deltas (raw bytes in bundle):

file	base	head	Δ
.build/www/_worker.js/lib/twins.js	0.92 KiB	0.97 KiB	+0.04 KiB (+4.5%)

Dictionary deltas (pd/)

144 → 144 deltas, 449.54 KiB → 449.97 KiB (+0.42 KiB).