[yangjj-iso]

Summary

• Fix false-positive Go CodeInterpreter.IsHealthy and waitRuntimeReady results when the runtime-check SSE response ends after only init or stdout. These executions have no error event, but also no confirmed completion; the existing checks incorrectly treated them as healthy/ready.
• Require a non-nil execution, no execution error, and a confirmed zero exit code. A completion reporting exit code 1 without a separate error event must also fail the probe.
• Preserve compatibility with older servers: RunCommand already infers exit code 0 from a successful execution_complete event that omits exit_code.
• Add HTTP regression coverage for incomplete probes, nonzero completion, and a second successful legacy-format probe. Empty responses remain rejected as before and are included as a control, not claimed as a newly fixed defect.

Testing

• New tests fail on unchanged production code: init-only, stdout-only, and nonzero completion wrongly pass health/readiness; polling stops after the first unfinished probe instead of waiting for the second confirmed probe.
• Go 1.20.14 final patch: CodeInterpreter tests passed 20 runs; full go test ./... -count=1 passed (12.272s).
• Final go test -race ./... -count=1 passed (13.425s); go vet ./... and git diff --check passed.
• An earlier full race-suite run hit the independent timeout-message diagnostic flake addressed by test(sdks/go): remove timeout-message race from connection diagnostic #2479. It was an assertion failure, not a race-detector report. That separate change is not included here; final runs above are against this PR's own two-file patch.
• Live sandbox/Jupyter and Redis integration tests not run. Tests use local HTTP/SSE responses and the real SDK parsing/readiness path.

Breaking Changes

• None: restores strict runtime-readiness semantics; successful old-server completion events still pass.

Checklist

• Clearly described motivation and reproduction
• Regression tests added
• Security impact considered: no auth, permission, transport, or retry changes
• Backward compatibility considered
• Documentation unchanged: a runtime probe without confirmed success should not report healthy.