CI overhaul (9/9): purge the agent-device runner-artifact cache before retry attempts #76

Part of the CI overhaul execution plan (#59, map #53). Order: 9 of 9 — blocked by 4/9 (#68) only to avoid churn in the same retry-shell files.

Decision: #56 research, flake 3 (docs/research/agent-device-ci-hardening.md on branch research/agent-device-ci-hardening): the restored runner artifact can fail with dyld: Library not loaded: libcurl.4.dylib because the cache key omits the macOS host build, and the rebuild fallback only fires on retriable health failures — so the dyld failure is a hard error today.

Scope

Local workaround: purge ~/.agent-device/apple-runner before each retry attempt in the device-contract job shell, so attempt 2 rebuilds the runner from scratch instead of restoring the same broken artifact. First attempts keep the cache (it's the fast path). The upstream report for the cache-key gap is filed separately (follow-up issue).

Files

• scripts/run-device-contract-job.mjs — in the exit-75 retry path, remove ~/.agent-device/apple-runner before launching attempt 2 (macOS only).

Ruleset / required-check changes

None.

Acceptance criteria

• Attempt 2 of an iOS device-contract job starts with no ~/.agent-device/apple-runner directory (visible in the attempt log).
• First attempts still restore/reuse the cached runner artifact.
• Exit protocol unchanged: contract failures still never retry.