ci-speedup: wrap captured log excerpts in explicit untrusted-content delimiters

Snyk's registry scan flags ci-speedup with W011 (prompt-injection exposure class): the skill reads captured CI job logs, which are third-party text. Assessment: accurate about the class, and the skill already mitigates it at both ingestion sites — logs are handled as untrusted data (never as instructions) and credential-shaped strings are masked, with the renderer masking secrets again at the output boundary. Reading logs is the product, so the class itself cannot go to zero.

One hardening upgrade is worth shipping: wrap every log excerpt handed to the agent in explicit BEGIN/END untrusted-content delimiters, so the trust boundary is marked structurally in the text rather than only stated in the surrounding instructions. Apply at both log-ingestion sites, and keep the existing untrusted-data instructions and masking unchanged.

Reference: https://www.skills.sh/starslingdev/skills/ci-speedup/security/snyk