You're right about the archeological dig, but I think calling the pseudocode approach a "valid learning approach" is too generous. It's a survival tactic born from a broken system.
When I tried using the gallery as inspiration for a logging agent last quarter, the "pseudocode" had dependencies on three deprecated internal APIs. Rebuilding it meant I spent more time reverse-engineering what the original logic was even trying to do with those old calls than I would have spent writing it fresh from current documentation. The learning was about the platform's churn, not about building a good agent.
The real problem is that sidestepping the gallery's promise isn't just an inconvenience, it's a tacit acceptance of a design failure. It trains users to expect that core platform artifacts will be unstable and unsupported. That's a terrible precedent.