0608 — Playwright covers the editor swap and the initial-JS ceiling

Context

The flows suite [0607] left two gaps. The CodeMirror mount had no test — the 2026-08-28 sandbox refactor checked it by hand. And website/CONVENTIONS.md ## Islands & loading promised a Playwright initial-JS ceiling that nothing enforced; the build-time import-graph walk that once enforced a ceiling was rejected on code-to-gain [0594], which named two obstacles for any revival: “initial” must be a moment (the sandbox starts booting PGlite in onMount), and a ~240-line walk is too much machinery for one number.

Decision

• Editor swap: a flow test waits for .cm-editor inside the first .sql-panel, asserts the static pre shell left .editor-area, then types into the editor and asserts the panel chip leaves “auto re-runs” — proof the swap wired setSql, not just that a div appeared.
• Initial JS: a ~20-line flow test sums script response bytes on the homepage at a 640px viewport — below the 761px sandbox gate the island never hydrates, so networkidle is a stable moment and the sum is the JS every phone reader pays. Raw (decompressed) bytes, measured 82,080; ceiling 96,000, so an island-sized chunk fails the build while small growth passes. A second assertion rejects any PGlite chunk request below the gate.
• The CONVENTIONS sentence now describes this check and keeps the import-graph walk rejected [0594].

Consequences

• Accepted blind spot: below the gate the sandbox island’s own chunk never loads, so a CodeMirror static import inside that island is invisible; covering it means counting on desktop before an unstable moment — the flakiness [0594] warned about.
• The ceiling is per-homepage; growth every page pays (a ClientRouter [0592], a footer island [0593]) still lands here, since the homepage loads those same shared chunks.