GCPSkillRegistry.search_skills crashes on the first catalog entry that fails Frontmatter validation

[doublenum]

🔴 Required Information

Describe the Bug:
GCPSkillRegistry.search_skills() constructs a models.Frontmatter for every hit returned by the Agent Registry skills:search endpoint with no per-item error handling. Frontmatter.name is validated against _SNAKE_OR_KEBAB_NAME_PATTERN (only a-z, 0-9, -, _), but real first-party catalog entries carry names outside that pattern — e.g. cloud.google.com-agent-platform-eval-flywheel (dots). The first such entry raises a pydantic ValidationError and sinks the entire search call, so search_skills is unusable against any catalog containing one non-conforming entry — including entries the caller never asked about.

google/adk/integrations/skill_registry/gcp_skill_registry.py, search_skills() (2.7.1, ~line 214):

results = []
for s in response_data.get("skills", []):
  results.append(
      models.Frontmatter(
          name=s.get("name", "").split("/")[-1],   # <- one bad name raises, whole call dies
          description=s.get("description", "") or "",
      )
  )
return results

Steps to Reproduce:

1. Install google-adk==2.7.1.
2. Point GCPSkillRegistry at an Agent Registry project/location whose catalog contains at least one skill whose last path segment does not match _SNAKE_OR_KEBAB_NAME_PATTERN (the public first-party entry cloud.google.com-agent-platform-eval-flywheel reproduces this).
3. await registry.search_skills(query="anything").
4. pydantic ValidationError for Frontmatter.name; no results are returned at all.

Expected Behavior:
A catalog entry that fails client-side frontmatter validation is skipped (optionally logged), and the remaining valid hits are returned. One malformed/non-conforming entry — which the searching agent has no control over — should not make skill discovery unusable for the whole catalog.

Observed Behavior:
pydantic_core._pydantic_core.ValidationError: 1 validation error for Frontmatter — name string does not match the snake/kebab pattern — raised out of search_skills(); the whole call fails.

Environment Details:

• ADK Library Version (pip show google-adk): 2.7.1
• Desktop OS: Linux
• Python Version (python -V): 3.12

Model Information:

• Are you using LiteLLM: No
• Which model is being used: N/A (client-side registry parsing; no model involved)

Suggested fix

Wrap the per-item Frontmatter construction in try/except, debug-log and skip the failing entry. We have been running exactly that as a client-side monkey-patch in production and it restores discovery against the real catalog:

for s in response_data.get("skills", []):
  try:
    results.append(
        models.Frontmatter(
            name=s.get("name", "").split("/")[-1],
            description=s.get("description", "") or "",
        )
    )
  except Exception:
    logger.debug("search_skills: skipped result failing frontmatter validation: %r", s.get("name"))

Happy to send a PR if that direction is acceptable.

Related context: since 2.7 get_skill() also validates names against the same pattern before building the resource URL (11101ac / #6805), so non-conforming catalog entries are now unloadable by design — which is fine, but they still appear in search responses, so search_skills needs to tolerate them rather than crash.

[a2105z]

Taking this right now, don't duplicate.