docs(website): screenshots for the seven pages the 0.66 audit flagged as image-less

[edspencer]

The v0.66 docs audit flagged seven pages that lean on describing a UI they never
show — the whole guides/ tree carried zero images, and getting-started.md,
the highest-traffic page, had none either. This adds one image to each, plus the
script that made them.

Follows #776, #777, #778, #779.

What is shown

Page	Shot
using/working-in-chats.md	the Adopt N native chats row above the chat list
using/working-in-chats.md	the Adopt native chats dialog, one row unticked so the button reads Adopt 2 chats
using/working-in-chats.md	the emerald Adopted badge on one row, beside three chats without it
using/creating-and-organizing-projects.md	Promote to project, name pre-filled from the chat
using/sending-files-and-images.md	a sent message — image thumbnail + file chip — which the page never showed; it only showed the staging tray
guides/agent-capabilities.md	the trigger tool picker with Bash ticked and its warning raised
getting-started.md	Home, immediately after npx @edspencer/paddock --here

Every image has real alt text describing what is in the frame.

Where they came from

A synthetic instance — fictional projects, fictional chat titles, /home/<demo>
paths, stock branding. No copy of anyone's real data was involved, so there is
nothing to scrub. Each frame was scanned at the DOM level before it was written
and then looked at.

The attachments shot needed state that did not exist: an attachment lives in the
attachment store, not in the transcript JSONL, so stage-attachments.mjs drives
the two real steps a browser takes (multipart upload, then chat:send carrying
the returned ids) rather than hand-writing a .jsonl and risking a shape the UI
only renders by accident.

Why the tooling is in the PR

A visual-design overhaul is in flight (#763–#767, #780–#783), so every one of
these will go stale at once. That is accepted deliberately: the prose is worth
illustrating now, and re-shooting is node capture.mjs rather than a person
re-deriving a dozen navigation paths. The script is sanitised for a public repo —
no absolute paths into one machine's clone, no private dev domain, base URL and
output from flags or env, and site-specific leak patterns in
$PADDOCK_LEAK_EXTRA. Hard-coding a private domain in the detection regex
would publish the string it guards; an earlier draft did exactly that.

Two framing bugs were fixed while re-shooting, both recorded in the script's
comments: an unframed element shot of a scrollable list is ~40% empty because a
list is as tall as its viewport rather than its content, and two unframed shots
of the same URL at the same viewport land as byte-identical files — which is
how one earlier capture appeared to be two.

Verification

• npm run build in website/ — clean, 51 pages, all seven images processed
  into dist/_astro/*.webp
• the sanitised script re-runs end to end and reproduces all seven shots
• caddyfile highlight warnings in the build log predate this branch (2 on the
  v0.66.2 tag, 2 on main)

Docs-only: no changeset, no version bump.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying paddock with    Cloudflare Pages

Latest commit:	ccad491
Status:	✅  Deploy successful!
Preview URL:	https://8a26ca74.paddock-7u2.pages.dev
Branch Preview URL:	https://docs-screenshots-0-66.paddock-7u2.pages.dev

View logs