* feat(ci): migrate Linux CI jobs to self-hosted runners on hanzo-k8s (#31)

Route all Linux CI jobs to self-hosted runners (hanzo-k8s label) on
our DOKS cluster instead of GitHub-hosted ubuntu-latest. This
eliminates queue wait times and runs jobs on dedicated infrastructure.

Migrated workflows: ci.yml (8 jobs), workflow-sanity.yml,
install-smoke.yml, formal-conformance.yml, labeler.yml,
auto-response.yml, stale.yml.

Kept on GitHub runners: Android (needs JDK/Gradle/SDK),
Windows (needs Windows OS), macOS (needs Xcode/Swift),
Docker release (needs Docker), npm release (needs registry).

Adjusted test parallelism: BOT_TEST_WORKERS=1 for 4GiB runner pods.

Runners: 3x github-runner + 2x github-runner-build pods in
actions-runner-system namespace, org-scoped for hanzoai.


* fix(ci): use python3 instead of python for self-hosted runner compat

The myoung34/github-runner:ubuntu-jammy image only has python3,
not the python symlink. Update the no-tabs check accordingly.


* fix(ci): bump runner memory to 8GiB and revert labeler to ubuntu-latest

- Scale self-hosted runners from 4GiB to 8GiB (node test OOM'd)
- Restore BOT_TEST_WORKERS=2 (fits in 8GiB)
- Revert labeler.yml to ubuntu-latest (GH_APP_PRIVATE_KEY org secret
  not available to self-hosted runners)


* fix(test): use path outside tmpdir for media root rejection test

On self-hosted runners the workspace is /tmp/runner/work/... which
falls under os.tmpdir() (/tmp) — an allowed media root. The test
was using process.cwd()/package.json which passed the root check,
letting the Discord token validation fire first with a different error.

Use /usr/share/ instead, which is never under an allowed root.


* fix(ci): reduce Node test heap to 2048MB to avoid OOM on self-hosted runners

2 workers × 3072 MB = 6 GB V8 heap on 8 GB runners leaves no room for
OS + Node RSS overhead, causing OOM kills during test cleanup. Reducing
to 2048 MB per worker (4 GB total) leaves enough headroom.


* fix(ci): reduce test workers to 1 per group to fit 12GB pod limit

test-parallel.mjs runs 3 vitest groups in parallel, each spawning
BOT_TEST_WORKERS processes. With 2 workers per group, total V8 heap
was 3×2×3072MB = 18GB, causing OOMKilled on 16GB runner nodes.
Reducing to 1 worker per group gives 3×3072MB = 9GB — fits in 12GB.


* feat(ci): switch to ARC (Actions Runner Controller) for auto-scaling

Replace manual myoung34/github-runner deployments with GitHub's official
ARC controller. Runner pods now scale on-demand (0→10) when jobs queue,
each getting its own 14Gi pod on the runner-pool node pool. Workers stay
at 2 per vitest group with 2048MB heap (3×2×2048 = 12GB, fits in 14Gi).


---------