Add fxci-task-cost-attribution skill

[jwmossmoz]

Summary

New skill that answers per-task and per-push FXCI cost questions across both clouds while the RELOPS-2330 production pipeline is still pending.

The workflow now uses a six-stage pipeline:

1. GCP per-task rows from the existing task_run_costs_v1 path
2. Azure per-VM-day cost rows from the Azure billing export
3. Azure vm-* task runs from FXCI task tables
4. Azure worker uptime from taskclusteretl.worker_metrics
5. Local DuckDB join for Azure per-task rows
6. DuckDB summaries across GCP + Azure

This is distinct from azure-cost-analysis, which is for pool/SKU-level Azure Cost Management rollups. This skill is for task, push, tree, kind, and label attribution.

Worker Metrics Update

Azure uptime now comes from generic-worker WORKER_METRICS lifecycle events in taskclusteretl.worker_metrics, rather than a first-task/last-task approximation from task_runs_v1.

That gives a better denominator without needing Azure Monitor. The docs still call out the remaining limitation: worker metrics are Taskcluster-side data and do not fully capture Azure VM lifetime after the worker has stopped but before the VM is deleted.

Public Data Cleanup

Removed the sample cost breakdowns and dollar figures from the public PR text and methodology docs. The skill documents the workflow and caveats without publishing internal cost totals.

Why Two Gcloud Accounts

Human IAM is split between moz-fx-data-shared-prod and moz-fx-data-billing-prod-9147. No single human principal can read both billing datasets, so the workflow runs the BigQuery halves separately and joins locally with DuckDB.

Structure

skills/fxci-task-cost-attribution/
├── SKILL.md
├── queries/
│   ├── README.md
│   ├── 01_gcp_per_task.sql
│   ├── 02_azure_vm_cost.sql
│   ├── 03_azure_task_runs_all.sql
│   ├── 04_azure_worker_metrics.sql
│   ├── 05_local_join.sql
│   └── 06_summary.sql
└── references/
    ├── README.md
    ├── auth-split.md
    ├── duckdb-local-join.md
    ├── methodology.md
    └── jira-context.md

Retires When

Once the RELOPS-2330 bigquery-etl pipeline lands and fxci_derived.task_run_costs_v1 covers Azure VMs end-to-end, most users can query that table directly. Until then, this skill is the local workaround.

Test Plan

• Confirmed taskclusteretl.worker_metrics has Azure Windows generic-worker lifecycle events
• Confirmed worker-metrics task durations line up with task_runs_v1
• Confirmed the new worker-uptime query shape returns worker-day uptime rows through Redash
• Confirmed public docs no longer include sample cost breakdowns or dollar figures
• Internal @references/* and queries/* links updated for the six-stage workflow
• No hardcoded user paths in skill files
• Reviewer to validate command lines on a Windows host