Regression test for #83 over-triggering on comment-only MODULE.bazel edit (bzlmod analog)

[tinder-maxwellelliott]

Summary

• Adds a passing regression-protection test for #83. The 2021 report was that a comment-only edit to WORKSPACE in a rules_go + Gazelle setup caused every external Go dep — and the go_binary depending on them — to be reported as impacted.
• The bzlmod-era analog of that bug is a comment-only edit to MODULE.bazel re-hashing the entire dep graph. This is exactly the over-triggering shape PR #314 was designed to prevent: rather than seed every target's hash with the raw MODULE.bazel content, bazel-diff parses bazel mod graph --output=json and only re-hashes targets when the resolved module graph differs.
• I verified by hand with the locally built CLI: prepending a comment to MODULE.bazel produces an empty impacted-targets list.
• New module_bazel_comment fixture: minimal bzlmod workspace with a single bazel_dep on rules_shell and two sh_library/sh_binary targets.
• This is a passing regression test (not @Ignored). If a future change goes back to seeding hashes with raw MODULE.bazel content, the test will fail. WORKSPACE mode is no longer supported in current Bazel, so a WORKSPACE-comment reproducer would need a legacy-Bazel fixture; the bzlmod equivalent is the right shape going forward.
• Test skips itself on Bazel < 8.6.0 (and on the 9.0.0 release that had the broken mod show_repo path), matching the convention used elsewhere.

Test plan

• bazel build //cli:cli-test-lib succeeds.
• bazel query '//...' against the new fixture returns //:lib and //:bin.
• bazel test //cli:E2ETest --test_filter=testModuleBazelCommentOnlyChangeDoesNotImpactTargets_regressionForIssue83 → PASSED in 2.7s on Bazel 9.1.
• Manual run with locally built CLI confirms a comment-only MODULE.bazel edit yields an empty impacted-targets list.

🤖 Generated with Claude Code