[Karanvir1729]

The runtime's measured visual checks (scroll locks left on, click blocks, invisible text, ads over content and the rest in src/visual/engine.ts) and their runtime fixes now reach the dashboard from production traffic, and the dashboard shows them. Before this, telemetry sent nothing about them, so a site full of stuck scrolls and frozen clicks looked clean on genclass.dev.

The owner approved the counts-only fields in VISUAL.md section 10 on 2026-10-27. Nothing else is sent: never a selector, element text, colour, URL, size or one event per finding.

Base: this PR contains PR 118

It is built on the live collector's deploy tree, 33df1e35: main merged with PR 118 (copy-prompts), with the conflicts resolved. That tree also has b0710d04, the demo estimate moved into stats(), which is on local main but was not pushed. "Fix them all" exists only in PR 118, and the collector deploy has to keep it. So this diff against main includes PR 118. Merging this PR lands both; merging PR 118 first shrinks this diff to its own three commits.

0.3.0-beta.4 belongs to PR 119 (the devtools pill, for the demo recording), so this targets 0.3.0-beta.5. Once PR 119 merges, main must be merged into this branch. Expect a conflict in the CHANGELOG's top section.

Runtime (packages/runtime)

• Session event: visual ("off", "local" or "on", the first-load stub's setting) and fix (the app turned runtime fixes on: visual: { on: true, fix }). Guard or heal is still needed, and the same event's effectiveMode says which applies. These two fields are the only first-load bytes.

• Summary event: once the checks have run, it carries:

  • counts.checks
  • counts.checksByRoute (at most 20 normalised paths)
  • counts.fixes { applied, undone }
  • runtime.visual { passes, ms, partial, tier }

  The checks' lazy chunk adds them (src/visual/bridge.ts withChecks: follow() wraps the telemetry client's push). First load carries no code for them, and a page where the chunk never loads sends summaries exactly as before.

• Console notice: printed once by the chunk when telemetry is on. The telemetry notice itself is unchanged.

• Docs: TELEMETRY.md, both READMEs, PRIVACY.md, VISUAL.md section 10 (approved, as built) and the CHANGELOG (Unreleased, 0.3.0-beta.5).

• Tests: test/telemetry-visual.test.ts (new); test/visual-bridge.test.ts (its stub gains fix). No review tests were edited, and no model-visible text changed (situation-v2.3).

Collector (telemetry-worker/src/dashboard.ts)

• Exact counters per page load from summaries, through the same high-water marks as detections: checks, check_kind, check_issue (kind TAB page), check_fix (applied, undone), check_loads, check_tier, check_partial and check_ms.
• Coverage counters from session events: visual_setting and visual_fix.
• Input filtering: only check-kind names and page paths are kept; junk keys are dropped.
• No migration. The counters tables hold any metric.
• stats() additions: series[].checks, checkSeries, previous.checks and totals.checksWithPage.

Dashboard (telemetry-worker/src/pages.ts)

• Their own kind, never a model finding:
  • a "Measured check" badge;
  • a card per check with what was measured, the pages, the usual fix, and a Copy prompt that measures the page again with genclass_scan;
  • a hatched slot in Fig. 1, in the muted ink rather than a new hue, so the validated palette is unchanged;
  • a hatched segment in the Pages table;
  • the check fixes as their own row in "What GenClass did";
  • their coverage in Health, including how to turn them on when no page load ran them.
• Counted as problems in the summary, the tiles, the rail, the app status (a check fix still in place counts as fixed live) and "Fix them all".
• Sample dashboard: shows simulated checks.
• Demo estimate: the Snow Day estimate now yields to the measured counts as soon as that project gets its first page load with measured checks (a check_loads row). Delete src/demo.ts after the recording.
• Tests:
  • test/dashboard.test.ts: the measured visual checks.
  • test/pages.test.ts: the measured visual checks. Existing expectations now count problems as detections plus checks, because the sample has checks; two tests about the model alone strip them with noChecks().

Not run here

Nothing ran locally. The Mac only edits files, the train VM is deallocated, and starting it needs the owner's OK. This PR's CI runs tsc, the build, both runtime vitest steps, the worker tests and the first-load size check.

First-load size: /auto had 11 B of headroom under bundle.test.ts (108.589 of 108.6 KB on main). The two session fields will probably go over it. If CI fails there, raising LIMIT_KB by the measured growth needs the owner's OK. The 0.6 KB growth limit in first-load-size.mjs should pass.

The worker has no tsc step and no tsconfig, so its TypeScript is not type-checked anywhere.

For the owner (nothing here is published, tagged or deployed)

1. Release: after PR 119's v0.3.0-beta.4, release this as v0.3.0-beta.5 through .github/workflows/release.yml. That means a version bump in packages/runtime/package.json and src/version.ts, the CHANGELOG heading, the exact install pins, then the v0.3.0-beta.5 tag. Then a worker PR like telemetry-worker: the newest runtime on npm is 0.3.0-beta.3; the act 3 rehearsal #114 for RUNTIME_ON_NPM and the CDN pins.

2. Collector deploy: from main once this PR (and with it PR 118's copy-prompts, which is live) is merged: cd telemetry-worker && npx --yes wrangler@4 deploy. Deploying plain main without PR 118 would remove the live Fix them all and the Connect checkboxes. No D1 migration comes with this PR. Still run d1 migrations list genclass-dashboard --remote first.

3. snowday-predictor.ca (Dizon's Cloudflare account):

   • pin @genclass/runtime@0.3.0-beta.5 in the script tag;
   • keep guard mode and data-visual="fix";
   • set the token to the "Snowday Predictor" project's key, gc_yTryldWvQz305xsmulOAmq. The live tag sends a key starting gc_Ypx0K, which is why that project has 33 page loads and no problems.

   Until it runs beta.5, the site sends no check counts (beta.4 doesn't), and that project's dashboard keeps the demo estimate.

Follow-ups (not in this PR)

• @genclass/mcp's genclass_analytics production summary does not mention the checks yet.
• The MCP view's tool table lists 8 tools and leaves out genclass_scan.

🤖 Generated with Claude Code