test: expand transcript payload smoke coverage

[nvphungdev]

Closes #2

Summary

• use sys.executable in the smoke harness so it works on systems that only provide python3
• add smoke coverage for nested transcript assistant payloads, malformed transcript lines, clarifying-question allow behavior, and caveat detection edge cases
• skip non-visible thinking / redacted_thinking transcript blocks when extracting assistant text for the stop gate

Verification

• python3 scripts/smoke_test.py
• git diff --check