I hope the release of 1.0 is an indication that Bazel is taking interface stability more seriously.
I have to say, having used a number of Google's tools over the years: basel, coral sdk, maps api and perhaps others, I'm always surprised at how 'how ya doing' the experience is. Don't get me wrong, I appreciate their tools and its openness, but I wish Google spend a bit more time doing software engineering, and less throwing shit against a wall ascertaining stickiness.
This is fixed now: https://github.com/tensorflow/tensorflow/commit/991aec351b57...
If you use Bazelisk (https://github.com/bazelbuild/bazelisk), you'll automatically get the matching Bazel version for TensorFlow in the future (and all other projects that have adopted the .bazelversion file).
Though, manually specifying a maximum Bazel version doesn't seem like it should be necessary. The authors of Bazel know what features they have changed since the version specified as the minimum, and Bazel itself knows what features the BUILD file uses. That information together could be used to calculate the maximum version.
Also, unless they had wildly different versioning pre 1.0, they could have just said "we are releasing 1.0". semver already states that 0.y.z should not be considered a stable API [1].