ForgeBook S4: Cut over to durable popularity projections after soak

[nicholasdbrady]

Parent

• coreai-microsoft/foundry-devx#173

Blocked by

• coreai-microsoft/foundry-devx#175
• Seven consecutive successful daily projection runs with current health state and zero reconciliation differences

Summary

Switch Forgebook’s build-time popularity reader from direct raw Application Insights queries to the durable Azure Table Storage projection owned by Foundry DevX. Preserve the freshly deployed page-view snapshot as the last-known-good fallback until the durable reader passes its soak and cutover gates.

User stories covered

• Rank the Popular carousel by production recipe page views in the trailing 30 days.
• Display production recipe page views accumulated since the June 2, 2026 launch.
• Exclude localhost, PR previews, non-recipe routes, and unpublished registry slugs.
• Keep deployments successful and counters stable when analytics data is stale or unavailable.
• Use GitHub OIDC and least-privilege Entra RBAC without storage keys or browser access.

Acceptance criteria

• The GitHub Actions identity has only the Azure Storage Table Data Reader access required for the aggregate table.
• The build-time reader uses Entra/OIDC authentication and reads the small materialized totals projection; no storage keys, SAS tokens, or public browser API are introduced.
• Only current registry.yaml slugs with a positive trailing-30-day view count are eligible for the carousel.
• The carousel is ordered by trailing30PageViews, capped at six, and each badge displays sinceLaunchPageViews.
• The reader validates metric version, asOfUtc, source coverage, and reconciliation health before replacing popular.json.
• Missing, stale, malformed, or unhealthy projection data causes the deploy to emit a warning and retain the committed last-known-good popular.json rather than failing or rendering an empty carousel.
• The generated snapshot remains deterministic and includes the as-of timestamp, launch boundary, ranking window, metric version, and both view measures.
• Reader tests use fixture projections and cover healthy data, unpublished slugs, zero recent views, stale data, unhealthy reconciliation, malformed rows, and unavailable Azure access.
• The existing homepage Playwright seam verifies trailing-30-day ordering and since-launch displayed views without hard-coding daily-changing counts.
• Registry validation and the production site build pass.
• Cutover occurs only after the Foundry DevX projection has completed seven consecutive successful daily runs with current health and zero reconciliation differences.

Testing

Test the external build-time behavior rather than Azure SDK call sequences. Use fixture projection entities at the reader seam, retain the existing homepage Playwright coverage, and record the seven-run production soak as a release gate.

Out of scope

• Provisioning or writing the aggregate store.
• Changing the canonical metric or Log Analytics summary rules.
• Giving the browser direct access to Azure Storage.
• Removing the committed fallback.
• Raw telemetry export, Logic Apps reporting, or proactive Action Group notifications.