fix(headless): classify incomplete terminal provider streams under Pier

[Astro-Han]

Summary

Post-merge external review of #1758 found that d783aa3a4's terminal-provider-stream classification only exists in the Harbor runner: a Pier trial whose last proxied provider request ends interrupted/failed/aborted (a 200 stream without its protocol terminal event = incomplete provider response) still reads the verifier reward and can record a graded model failure where Harbor throws infra. This ports the semantic to the Pier runner.

• Extracts the predicate as incompleteTerminalProviderRequest in harbor-task-runner.ts (the module both runners already import shared helpers from); Harbor's behavior is byte-identical
• pier-task-runner.ts throws PierInfraError ('infra_failed', telemetry artifact attached) at the same point in its flow — after the exit-code check, before reading the reward
• Both DeepSWE arms run under Pier, so paired comparability was never broken; this fixes the classification itself for every Pier benchmark

Verification

• RED first: new pier-runner test fails with "Missing expected rejection" before the fix
• The test drives the runner's real seam — a hub lease whose recorded telemetry ends interrupted must throw, completed must take the reward path (the proxy's recording of truncated streams is already covered by the Harbor live-SSE test on the shared proxy machinery)
• packages/headless suite: 1456 pass / 0 fail; format/lint clean

Review focus

The timed-out-trial exemption is shared with Harbor: complete timed-out trials settle by deadline, so their truncated tail request is expected.