Avoid noisy failed trace span for expected not-found SKILL.md during lazy skill probe (fixes #3889)

[MuhammadUsman-Khan]

Summary

Avoid noisy failed trace span for expected not-found SKILL.md during lazy skill probe.

Fixes #3889

Root cause: LocalDirLazySkillSource.load_skill used the instrumented SandboxSession.read() to probe for SKILL.md. On first lazy load this probe raises WorkspaceReadNotFoundError (expected) and the instrumentation recorded a failed child span even though the exception is handled and the overall load succeeds.

Solution: For the existence probe only, unwrap a SandboxSession wrapper (if present) to get the underlying BaseSandboxSession and call its read() directly. This prevents the expected handled-not-found from being recorded as an error span. The change is tiny and contained to src/agents/sandbox/capabilities/skills.py.

Files changed:

• src/agents/sandbox/capabilities/skills.py
• tests/test_lazy_skill_probe_unwrap.py (new)

Testing performed:

• Added unit test to ensure load_skill proceeds when the inner session read raises FileNotFoundError.
• Recommend running the issue reporter's repro to verify traces no longer show the error span.

Limitations / follow-up:

• The defensive unwrap relies on the wrapper exposing _inner and matching the wrapper type name/module. This matches the approach used in other parts of the repo (session sinks).
• If other probes or reads should avoid instrumentation, apply the same pattern selectively.

[seratch]

Thanks for sharing this idea. The issue was resolved by #3893