Observability built as a first-class surface rather than a debug page, mirroring the pipeline's own stats.py "trust dial". No chart library — every mark here is a div grid, an SVG path or a stroke-dasharray, which is why the whole visual language stays inside the token system.
Six numbers that answer "is the Well healthy" before you read anything else. Each tile carries a label, the value, and a subtitle that names the underlying table or gives the comparison — a number with no denominator is not a signal.
entities subtitle interpolated snap.totals.edges raw while the value above it went through fmtNum (so "1,868" sat beside "1634 typed edges"), and the tokens subtitle claimed an "embedded substrate" while the coverage ring beside it correctly read 0.0%. Subtitles now run through the same formatter as their values, and the tokens subtitle is computed from embedCoverage at the ring's own precision. Both are logged as fixed on the issues page.Four mark types, each chosen for a specific question. Bars compare categories. The sparkline shows throughput shape over 14 months. Rings show a proportion you are meant to feel bad about when it is low. The depth column is the only one that also encodes the product's own depth-is-age rule.
vector-effect: non-scaling-stroke, because the SVG is stretched non-uniformly by preserveAspectRatio="none".Every chart lives in the same shell: a card with a two-part head — a title stating what is measured, and a mono note stating how. The note is where the honesty lives: "drives RLS, not just display", "% of requests where the lens was injected", "atoms by year · older sits deeper".
measuring the Well…. No skeleton shapes — a skeleton implies the layout is known, and here it genuinely is not until the snapshot arrives.measuring the Well…