Overview fidelity follow-up

The deployed overview diverged from the accepted V11/V12 concept: heavier numbers, a separate game-time line chart, a permanently narrow atlas, and oversized token/finance sections. This follow-up changes the production React overview, not just the local HTML.

Presentation

• The mock's embedded Inter 400 and 600 files have identical SHA-256 hashes to the installed font assets. Its 500-weight requests fall back to 400 because it does not embed 500. Use explicit regular weights for overview numbers and headings; preserve the Users page.
• Game time replaces the KPI detail surface with Interest atlas. The Worlds shortcut selects that same view. No second atlas below retention. Sources remains reachable for each selected metric.
• Atlas restores the discovery strip, roomy genre chips, Any/All matching, language and content filters, full-width cover map, and sortable List view. Both includes unknown content classifications; Unknown remains separately selectable. Limited maps to all; Limitless maps to sensitive, r18, and r18g.
• Map defaults to 24 worlds, expandable to 80; covers use area proportional to players. Two strongest audience edges per world are shown by default, with the selected world's mapped connections on selection. Same-family links are dashed. Shared-player neighborhoods use the concept's weighted modularity merge on the bounded displayed graph; tag labels describe membership, never determine it. Counts are editions, not summed unique audiences.
• World details appear below the full-width map after selection. All five measurements remain available. The live report's D7 return definition is retained and labeled D7, not the concept's different W1 window. Unrecorded historical time remains unavailable.
• Phones default to the ranked list and can switch to the map. Genre and content controls wrap; selection exposes readable details rather than shrinking all map labels.
• Token trend height is 125px. Flow defaults to a bounded 390px presentation with Expand available; all funding origins remain present. Model volume initially shows five rows.
• Finance retains two primary cash totals, reduces trend height to 185px, and compresses readout and unit-economics spacing. Receipts and provider-reported costs retain their existing definitions.

Cached periods

Prepared insight responses remain in the existing per-admin QueryClient for 30 minutes after becoming unused. Freshness remains 30 seconds, with visible observers polling and SSE revisions invalidating active reports. Returning to a cached period renders that period's saved response synchronously, then revalidates when stale. Old-period data is never presented under a newly selected period.

After the first overview response, a delayed task warms overview and finance ranges, plus the selected token/world detail when applicable. Warming issues batches of at most two requests, skips existing cache data, and stops starting new batches after navigation. Pointer/focus intent can warm the next view. These are authenticated reads of prepared report endpoints; no new database aggregates, writes, or worker jobs are triggered. Existing Users caching and authorization remain unchanged.

Reference and checks

The local yumina-admin-concept-v13.html preserves V12 and changes only the finance density; its builder and generated file live in the ignored money-dive directory. All inline scripts parse.

Twenty focused admin tests pass, covering navigation, Users cache/rank contracts, period cache reuse and isolation, bounded prefetch, and audience groups independent of genre labels. Root typecheck and production build are run for the final source. No browser-based visual verification was performed: repository instructions prohibit browser use without an explicit request. Visual decisions were checked against the supplied screenshots and the actual concept CSS/JavaScript; pixel-for-pixel parity is not claimed.

Live verification

Released commit 22308367e5c0b8c242384d4d078ec711552fa2f6 as Railway deployment 2abdc1ac-d767-4c56-9858-d4e94c8015fd. At 02:39 UTC, both running instances reported healthy primary DB, read DB, and Redis connections. Each could read all 21 prepared reports; the slowest individual Redis reads were 37.02ms and 17.10ms in these checks. Those figures are internal snapshot reads, not browser latency measurements.

Overview, Users, and Moderation public HTML served the same new /assets/main-BoN0DPvL.js entry with no-store headers. Admin data endpoints still returned 401 anonymously. The actual admin overview JS/CSS assets contain the new atlas, neighborhood controls, prefetch logic, compact flow, and regular metric weights. Runtime and analytics-worker deployments were not replaced by this frontend-only change.

A bounded synthetic Node check measured layout plus neighborhood grouping at 3.64ms for 24 worlds / 84 edges, and 10.87ms for 80 worlds / 280 edges. This is a local algorithm check, not a mobile rendering benchmark.