Merge template_usage_stats *_mins into app_usage_mins JSONB

[EhabY]

Phase 4 of the Scalable Approach for Adding New IDE Session Types RFC. The durable template_usage_stats table still has dedicated vscode_mins/jetbrains_mins/ssh_mins/reconnecting_pty_mins columns, so adding an IDE would still require a schema change there; this phase makes the durable rollup layer dynamic too.

Scope (RFC: "template_usage_stats must also be migrated"):

• UpsertTemplateUsageStats: have the agent_stats_buckets CTE produce app_name/app_minutes entries that flow into the existing app_usage_mins JSONB, the same way web app usage is stored today
• Key scheme: builtin session names will share app_usage_mins with web app slugs; confirm how the insights queries tell them apart before writing the backfill
• One-time migration: backfill the dedicated *_mins values into app_usage_mins for existing rows, then drop the dedicated columns, including the always-zero sftp_mins (dead since its TODOs were never implementable)
• GetTemplateInsights: read IDE minutes from app_usage_mins instead of dedicated columns; convertTemplateInsightsApps builds the builtin apps list dynamically
• GetTemplateAppInsights stays separate (it filters via workspace_apps, which builtin IDE sessions never join)

After this phase, a new IDE flows agent → normalized table → JSONB rollup with zero schema changes, which is the point of the RFC.

Done when

• Rollup writes IDE minutes into app_usage_mins and the dedicated columns are dropped
• Backfilled historical data produces identical GetTemplateInsights results
• A new app name reported through phase 1 shows up in template insights with no schema change

Depends on #27410.

---

🤖 Created by Coder Agents on behalf of @EhabY.

[linear-code]

ENG-3077

[EhabY]

Two decisions and one measurement from implementing phase 1 (#28124 → #28127).

Per-IDE history starts with this phase. DeleteOldWorkspaceAgentStats purges workspace_agent_stats about a day past the rollup watermark, and #28127 has the rollup store family minutes rather than per-app minutes. No backfill can recover per-IDE detail for any period before this phase ships. Family-level history stays correct throughout, so nothing the current surfaces expose is lost.

Use a new session_mins jsonb column, not app_usage_mins. The scope note asks to confirm how the insights queries tell the two key kinds apart, and as specified they cannot: app_usage_mins is keyed by workspace_apps.slug, joined via AND tus.app_usage_mins ? apps.slug and read as (tus.app_usage_mins -> apps.slug)::smallint (insights.sql:266 and :254). An IDE key named vscode would be attributed to any template with a web app slugged vscode, double-counting it in GetTemplateAppInsights. zed already appears as an app slug in this repo's tests. A separate column removes the ambiguity and leaves GetTemplateAppInsights untouched.

Backfill cost. On a synthetic 8M-row template_usage_stats (1301 MB, PG13), the backfill UPDATE takes about 67s holding ACCESS EXCLUSIVE, roughly 8s per million rows and scaling with retained history. Dropping the five *_mins columns is metadata-only, about 2ms. That argues for a separate migration rather than folding this into 000569, which completes in a few seconds.

Not required for correctness. With read-side family grouping in #28127, a new IDE already counts in its family with no schema change. This phase adds per-IDE granularity, so it can be scheduled on its own merits.