mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-06 03:38:23 +00:00
Claude
bb239c51ed
§8 of the interop plan ends with a requirement: the metadata-exposure analysis "should be surfaced in the UI if we ship this, not buried. A Marmot group and a cordn group have materially different metadata exposure and users cannot infer that from either one looking like a group chat." GroupExposure existed so there would be something to render; this renders it. The disclosure sits where it can still change a decision — on a pasted cordn1… link, before joining, rather than in a settings page nobody opens. The screen reads the link, names the coordinator and its relays, and shows the exposure. It deliberately does not join: joining needs a live coordinator and Tier B is blocked on licensing (§7), so a join button would be half a feature whose other half has never been run. - commonsUI/.../cordn/ui/: CordnExposureCard, CoordinatorHealthRow, CordnGroupBadge. Material3 only, so Desktop gets them for free. - commons/.../CordnLinkInspection: parse-and-disclose, headless and tested, so the screen only draws. A ref naming no coordinator is valid but not followable, and carries no exposure — inventing a threat model for a server we cannot identify would be worse than saying nothing. - Settings → Cordn group link, with a route and search keywords. Rendering it found two bugs compiling could not: - The severity colours were not monotonic. `tertiary` for the middle level rendered pink against a dark `onSurface` for the worst one, so "under a throwaway key" read as more alarming than "tied to your real account". Emphasis for the worst case is now weight rather than another hue. Nothing uses `error`: everything on the card is how cordn works, not a fault, and painting normal operation red teaches people to ignore red. - Note order is part of the disclosure. Cross-group linkage (§8.2), the item nobody predicts, sat below message padding, the least consequential one. notes() now returns most-surprising-first. CordnExposureRenderTest rasterises the surface headlessly (ImageComposeScene, software Skia, no display, no new dependency — the same Compose Desktop artifact jvmMain already uses) and asserts on pixels. It catches what a compile cannot: a missing string resource, which throws at render because the generated Res.string.* accessors compile regardless, and a composable that draws nothing. Mutation-checked — hardcoding the card background and disabling the §8.2 conditional each kill exactly one test. The first version of the theme test sampled only the page background and let the hardcoded card through, so it now samples inside the card as well. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012BfD4txdnsaPRXmNXbup9n