You're right, in the main repository it does. I only looked at my clone and didn't recognize that recomputations are actually triggered by *my* changes. So all is good, sorry for the noise.

If a current cached estimate is available, doesn't the heuristic catch this and simply return the cached value without triggering a recomputation?

In the eager search code, we currently reevaluate a lazy evaluator h for a state s if h has a cached estimate for s. I think this does more work than necessary though: if the cached estimate is not dirty, we don't need to recompute h(s).

I propose to add a method Evaluator::is_cached_estimate_dirty() and to recompute h(s) only if is_estimate_cached(s) and is_cached_estimate_dirty(s) both return true.