Fix #259 / #227 .bzl macro changes now invalidate targets

[tinder-maxwellelliott]

Summary

Fixes #259 and the duplicate #227.

Root cause. Bazel pre-7 populated Build.Rule.skylark_environment_hash_code in the query proto, so any change to a .bzl file loaded by a rule's BUILD file naturally bubbled into that rule's hash. Bazel 7+ leaves that field empty, so an edit like adding print() to a macro body no longer invalidated any caller — the emitted rule attrs were byte-identical.

Fix. BuildGraphHasher now walks every queried SourceFile target's subincludeList (the Build.SourceFile.subinclude proto field, which already lists every .bzl the BUILD file loaded), softDigests each main-repo .bzl, and mixes the union of digests into the workspace-wide seed.

Key properties:

• Conditional seed mix-in. Only mixed in when at least one main-repo .bzl was actually hashed, so the seed (and every target's hash) is byte-for-byte stable for workspaces that don't load any tracked .bzl files. Cached hashes from before this change are preserved for those users.
• External .bzl files are ignored (@repo//..., @@canonical//...) because SourceFileHasher.softDigest returns null for non-main-repo labels. This keeps the seed stable across BCR fetches that don't actually change repo contents.
• Intentionally conservative. A single .bzl edit re-hashes every target. .bzl edits are rare; silently missing them is the worse failure mode. Per-package precision would require mapping each rule to its package's BUILD file, which isn't a direct dep relationship in the query proto.

Also updates the macro_invalidation reproducer to use a print()-only mutation (the exact example from the issue) — the original cmd = mutation already propagated via rule attrs and wasn't the real bug. @Ignore removed; test renamed to a regression test.

Test plan

• bazel build //cli:bazel-diff //cli:cli-test-lib succeeds.
• bazel test //cli:BuildGraphHasherTest → PASSED (verifies the conditional seed mix-in doesn't shift hashes for workspaces with no tracked .bzl files).
• bazel test //cli:E2ETest --test_filter=testMacroBzlChangeImpactsCallers_regressionForIssue259And227 → PASSED in 15.6s (the reproducer that used to fail now succeeds).
• Broad non-Maven E2E sweep (testE2E$|testE2EWithNoKeepGoing|testMacroBzlChangeImpactsCallers|testBzlmodLocalPathOverride|testModuleBazelComment|testGenerateHashesWithCquery|testFineGrainedHashBzlMod|testBzlmodShowRepo|testExcludeExternalTargets|testCquery|testDetermineBazelVersion|testE2EIncludingTargetType|testE2EWithTargetType) → PASSED in 123s. The four Maven-fixture tests (testFineGrainedHashExternalRepo, testUseCqueryWithExternalDependencyChange, testUseCqueryWithAndroidCodeChange, testBzlmodTransitiveDepsCquery) fail locally with Cannot find Java binary bin/java from coursier — environmental (the bazel test sandbox doesn't propagate JAVA_HOME to the nested bazel invocations that run coursier). CI's setup-java step passes them.
• Manual repro confirmed end-to-end with the locally built CLI: adding print("miniature: " + name) to miniature.bzl produces //:logo_miniature in the impacted-targets list (previously it produced an empty list).

🤖 Generated with Claude Code