ci: opt this repo's docs build into the GPU upgrade pass (#885 part 2)

[ocots]

Phase F — the last piece of #885. Everything it depends on is merged: CTActions#71 (the reusable build-gpu job), #941 + #945 (gpu.md executable, CTBase 0.30.4-beta).

Full design and the feasibility probe behind it: .reports/campaign/P-gpu-docs-executable.md.

What this wires

with:
  use_ct_registry: true
  gpu_runner: '["occidata"]'
  gpu_timeout_minutes: 90

• gpu_runner opts this repo into the reusable's build-gpu job. The existing GitHub-hosted build job is untouched and still deploys first, every push and tag — build-gpu runs only once build has already published (needs: build) and, with continue-on-error: true, can never fail the workflow. On success it redeploys the same site with real GPU output; on failure, or if occidata is queued past its budget, the site build already published is simply left in place. Docs publishing is never on the critical path of a self-hosted box. build-gpu's own if restricts it to push/tag, so PR-preview builds are unaffected regardless of what's passed here.
• gpu_timeout_minutes: 90 — real margin over the 2265s (~38 min) measured on occidata during the Phase D probe (2026-09-02), not the reusable's blanket 120 default. Budgets for the slower case right after occidata-runner-maintenance.yml's Monday 02:30 UTC cache purge, not the same-day warm case that was actually measured.

New: a concurrency group

concurrency:
  group: documentation-${{ github.ref }}
  cancel-in-progress: true

Load-bearing now that a GPU pass can run behind a push for up to 90 minutes: without it, two pushes to main in quick succession could leave an older run's GPU pass finishing — and redeploying — after a newer run's CPU build has already published, showing older content over newer. cancel-in-progress cancels the whole older run, its queued or running GPU job included. Tags carry distinct refs (the tag name, not main), so a release build is never cancelled by an unrelated push to main, and vice versa.

What this PR's own CI will and won't show

build-gpu only fires on push/tag, by design (see above) — so this PR's own preview build stays exactly as fast as every other docs PR today, and does not exercise the new job. The real test is the first push to main after merge, where build deploys as always and build-gpu should show up as a second job in the run graph, landing on occidata behind it.

🤖 Generated with Claude Code