feat: normalize workspace agent session counts into a child table - #27952

Supersedes #27179 (same branch, squashed and rebased).

Phase 1 of the Scalable Approach for Adding New IDE Session Types RFC, tracked in #27410.

Summary

Replaces the four fixed session_count_* columns on workspace_agent_stats with a sparse session_counts JSONB object keyed by normalized app name. New IDEs can report session counts without schema changes, while existing APIs and metrics continue to expose the four well-known families.

Problem

Session counts were pinned to four hardcoded columns (vscode, jetbrains, reconnecting_pty, and ssh). Every new IDE was mislabeled or dropped, and adding a name required changes across the schema and application.

A normalized child-table implementation was benchmarked at up to 10 million rows and caused substantial read, write, storage, WAL, and migration regressions. Keeping the normalized object on the parent row avoids the extra relation and join overhead.

Fix

• Agent API v2.11 adds a session_counts map to Stats, deprecating fields 8 to 11. The server converts the fixed fields only when the map is empty.
• coderd/idemetadata holds the shared vocabulary. App names are normalized at ingestion, reports are capped at 64 distinct names, and overflow is combined into unknown.
• workspace_agent_stats.session_counts stores one sparse JSONB object per stats row. Arbitrary keys are preserved, while existing read paths extract the four well-known names.
• Agent counters are created on demand. POST /workspaces/{workspace}/usage and coder ssh --usage-app accept arbitrary app names.
• Usage queries process the latest completed minute without including rows before the requested cutoff, and template insights deduplicate activity by template, user, and minute.
• Removes agentsdk.Stats, which has been unreachable since the HTTP stats endpoint was removed.

Migration requirement

NOTE: Migration 569 takes an ACCESS EXCLUSIVE lock on workspace_agent_stats, backfills the retained raw-stat window, rebuilds the template-insights covering index, and drops the four fixed columns.

Deployments with retained rows containing positive session activity must have a template_usage_stats rollup from within the previous 24 hours. If the watermark is missing or stale, the migration aborts before changing the schema. Run the previous Coder version until template usage stats roll up, then retry the upgrade. Check coderd logs for failed to rollup data if the watermark does not advance.

Idle-only databases can upgrade without a watermark because their rows already match the {} default. Databases containing only raw stats older than the 180-day retention fallback can also upgrade.

With a healthy watermark, the 10 million-row benchmark converted approximately 333,000 rows in 4.44 seconds on PostgreSQL 17 tmpfs and 6.13 seconds on durable storage. Reads and writes remain blocked for most of the migration, and runtime scales with recent ingestion volume.

Mixed-version note: a new CLI passing a custom --usage-app value to an older coderd gets a 400 because old servers still validate against the fixed allowlist.

Follow-ups: #27411 (expose per-app counts in stats, Prometheus, UI), #27412 (connection_logs.type ENUM to TEXT), #27413 (template_usage_stats *_mins to JSONB), coder/vscode-coder#1044 (report the real IDE from the extension).

PR description updated by Coder Agents.