Behavior of Bazel's --experimental_reduce_costs flag for low‑traffic remote HTTP cache usage

Determine whether enabling Bazel’s --experimental_reduce_costs flag reliably lowers bandwidth usage for low‑traffic builds that rely on a remote HTTP cache without introducing correctness risks.

The flag is documented to skip uploading certain large outputs, but the criteria for what is omitted and how Bazel guarantees that dependent actions still receive needed data are not fully specified. Unclear factors include the stability of action keys, the possibility of stale cache entries when downstream actions depend on omitted outputs, and the traffic volume at which the bandwidth savings outweigh any potential recomputation costs.

• What conditions guarantee that skipped uploads do not cause cache misses for dependent actions?
• How does Bazel decide which outputs are safe to omit under --experimental_reduce_costs?
• At what approximate request frequency does the flag produce net cost savings for a remote HTTP cache?