recon(sprint-u0): web dashboard census, colour/contrast/streaming/focus findings

[konjoinfinity]

Summary

Read-only reconnaissance sprint (Sprint U-0 v2) on the axum-served / SvelteKit web dashboard, emphasis on the Loop Stacks page (/stacks) and its interactive controls. This is evidence for a design decision, not the decision itself, and no line of CSS/HTML/JS/Rust serving the dashboard was modified.

• Pre-flight passed: ref 043ca18 (HEAD, forked from main post-merge of PR fix(web): xN loop-count grammar, facet colors, and running-card chrome #181), all three canary selectors (.pop.sched, .chipinput, .cfgrow) resolved live before any capture work began.
• tools/recon/: a self-contained Playwright fixture/capture harness (own package.json, own node_modules gitignored) that mocks the app's REST + WebSocket surface deterministically (fixed ids/timestamps, no PRNG), reaching all 13 fixture states via the same composer/run-stack UI interactions a real user takes.
• recon/: the full census output, 55 page-state screenshots (11 reachable states x 5 viewports, S7/S8 correctly absent, no web consumer exists for budget-tier or cache-hit-ratio degradation), 8 labelled component state grids, a 20-frame motion strip, two 30s videos, and REPORT.md with all six censuses (colour/contrast, streaming path, layout stability, animation inventory, component matrix, keyboard/focus/layering).

Headline findings

1. Streaming path (the sprint's stated open question): the CLI is invoked in genuine incremental-streaming mode and read line-by-line, but claude_events.rs's parse_stream_event explicitly discards every content_block_delta before it reaches the UI. Text arrives as one atomic chunk per complete assistant turn, not incrementally. This is a lopi-agent plumbing fix, not a lopi-ui CSS fix.
2. Keyboard focus: only 5 of 25 sampled tab stops (20%) have any visible focus indicator. Every icon button in the stack-controls dock toolbar and every quick-insert chip token focuses invisibly (root cause: a transparent box-shadow on focus). Popovers also don't trap focus or restore it to the trigger on close.
3. Contrast: two real WCAG AA failures at ~1.1:1 (#231000 on near-black), not a subjective judgement call.
4. Animation: already at the sprint's own stated target (1 concurrent CSS animation during a running prompt), but a 250ms-interval JS timer per running card (agents.ts) is a real, quantified motion source the CSS-only audit alone would have missed.

Options (not recommendations) and specific open questions for each area are in REPORT.md sections 10-11.

Test plan

• Pre-flight canary resolution confirmed live in DOM before capturing
• git diff --stat against RECON_REF touches only tools/recon/, recon/ (verified, 110 files, all under those two paths)
• All censuses backed by live computed-style/DOM extraction or checked-in real capture data (artifacts/STREAM_CAPTURE.jsonl), not source-code inference alone
• Human review of REPORT.md sections 10 (options) and 11 (questions) to scope Sprint U

---

Generated by Claude Code