feat(observability): consolidate the 3 AI-usage dashboards into one, filterable by provider/feature/model

[JSONbored]

Summary

Phase B2 of the Grafana/AI-usage observability audit (follow-up to #5247/#5259, which fixed the
underlying ai_usage_events instrumentation this dashboard depends on).

• claude-usage.json, codex-usage.json, and orb-ai-usage.json overlapped heavily — the same
  ai_usage_events/Prometheus queries sliced 3 different ways (confirmed one exact duplicate: the
  "Cumulative AI cost by provider" panel existed byte-identical in both claude-usage.json and the
  infra dashboard gittensory.json).
• orb-ai-usage.json's SQL panels had no time-window scoping at all — every panel was a silent
  all-time aggregate regardless of what the dashboard's time picker showed.
• No provider/feature/model filtering existed anywhere, despite ai_usage_events already carrying
  all three columns.

Replaces all 3 with grafana/dashboards/ai-usage.json:

• Dynamic $provider/$feature/$model template variables (query-backed against
  ai_usage_events, never hardcoded — a self-hoster's own configured providers/features/models show
  up automatically, matching the existing $repo variable convention from the maintainer dashboard).
• Every durable-log panel now properly time-scoped (unixepoch(created_at) >= $__from ... < $__to).
• New "tokens by feature over time" and "cost share by feature" panels so an operator can see what's
  using the most tokens and spot spikes at a glance.
• A "live Prometheus counters" section (real-time signal — the durable log lags a completed call).
• Claude Code's own OTEL session telemetry kept in its own clearly-labeled, non-joinable section
  (structurally different data: session-scoped, no repo/PR attribution) rather than force-merged.

Also removes the now-duplicate "AI Usage & Cost" row from gittensory.json's infra dashboard, and
updates resource-hub.json's dashboard catalog (which was also missing gpu-metrics.json entirely).

Test plan

• npm run typecheck — clean
• node scripts/validate-observability-configs.mjs — dashboards valid
• npm run test:ci — full local gate green
• New test/unit/selfhost-grafana-ai-usage-dashboard.test.ts: template-variable shape, every
  panel scoped by provider/feature/model + time window (except the deliberately-unscoped
  "missing usage" panel, which has its own dedicated test), byte-for-byte carried-over Prometheus
  exprs, real sqlite3-CLI empirical narrowing tests (provider/feature/model filters actually
  narrow counts; "All" still sums everything; the missing-usage panel finds NULL-provider rows
  regardless of which provider/model they'd have had).
• Updated test/unit/selfhost-grafana-dashboard.test.ts: regression test proving the duplicated
  AI panels are gone from gittensory.json.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.37%. Comparing base (864c52d) to head (f2a5ccd).
⚠️ Report is 2 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5267   +/-   ##
=======================================
  Coverage   94.37%   94.37%           
=======================================
  Files         474      474           
  Lines       40114    40114           
  Branches    14627    14627           
=======================================
  Hits        37856    37856           
  Misses       1583     1583           
  Partials      675      675           

Flag	Coverage Δ
shard-1	46.27% <ø> (-0.15%)	⬇️
shard-2	34.73% <ø> (+0.01%)	⬆️
shard-3	32.05% <ø> (-0.05%)	⬇️
shard-4	31.82% <ø> (-0.26%)	⬇️
shard-5	33.59% <ø> (-0.07%)	⬇️
shard-6	45.06% <ø> (+0.30%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.