[tinder-maxwellelliott]

Summary

Reproducer for a report against a precomputed impacted-targets file: its labels named .../LibraryTests packages that the consumer's checkout no longer had, so the consumer's bazel query --query_file aborted with no such package and discarded all ~33k labels. The suspected mechanism was that bazel-diff counts a target present only in the "from" hashes as impacted.

The reproducer does not reproduce the bug. Every path walks the "to" hashes only, so a from-only label cannot come out of the diff itself. The new cases pin that, which narrows the report down to the hashes the diff was handed: the "to" hash file must already have contained the deleted packages, i.e. the generating checkout still had them on disk.

What changed

• New fixture tests/resources/workspaces/deleted_package: //keep, a leaf //keep/LibraryTests:keep_tests package depending on it, and //consumer depending on it. The tests delete the whole LibraryTests package and edit keep.sh, so the diff is non-empty and the deleted package would have been impacted had it survived.
• regressions::deleted_package_labels_are_not_impacted: get-impacted-targets with and without --depEdgesFile. Asserts the from-hashes contain the deleted target and the to-hashes do not, that the impacted set is exactly //keep:keep, //keep:keep.sh, //consumer:consumer, and that bazel query --query_file over that set succeeds at the "to" checkout, the way a CI consumer uses the file.
• regressions::serve_deleted_package_labels_are_not_impacted: the same through serve in the default mode and with --partialGraph, where the deletion must force a full generation (fallback reason names the deleted path) rather than a derivation from a base that still contains the package.
• tools/e2e/rust_e2e_cases.bzl regenerated (make regen-e2e).

Both cases pass locally under Bazel 8.8.2 (~25s each).

For the reporter

Two mechanisms on the generating side fit the symptom and are worth checking before anything changes here:

• The precompute generated "to" at a different tree than the consumer's commit (merge commit, stale base).
• The generating working tree kept files git does not track. serve switches revisions with git checkout --force, which removes tracked files deleted between revisions but leaves untracked and ignored ones in place. Generated, gitignored LibraryTests BUILD files left over from an earlier generator run would stay visible to bazel query at every revision.

Diffing the failing artifact's labels against the generating host's hash file for the "to" commit tells which.

Follow-up: the serve working-tree check for the second mechanism is in a separate PR.

🤖 Generated with Claude Code