Hosted gem rollback / remove still strip the DEPENDENCIES ! when the user's source … do block holds an assignment-form v = if/case/begin … end, so frozen installs fail on Bundler 4.0.19+ #1476

[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).

Summary

#1056 is closed (fixed by #1303), but one shape still breaks. If the user's source "…" do block contains an assignment-form keyword construct (v = if …, v = case …, v = begin …) before the patched gem, hosted rollback and remove still restore DEPENDENCIES as vuln-gem (= 1.0.0) without the !. Bundler pins a gem declared in a source block with !, so the restored lock no longer matches the Gemfile. On Bundler 4.0.19 and later, every frozen or deployment install then fails with exit 16. The unwind reports success (exit 0).

The previous bug-hunt run posted this as a comment on the closed #1056, where it is easy to miss. This issue tracks it on its own.

Impact

After socket-patch rollback or remove, CI and deploy checkouts that use BUNDLE_FROZEN / BUNDLE_DEPLOYMENT fail with "Your lockfile needs to be updated, but it can't be because frozen mode is set". The restored Gemfile also comes back dedented (gem "vuln-gem", "1.0.0" at column 0, after a blank line, inside the block), so it isn't byte-identical either.

Repro

I used a scratch copy of crates/socket-patch-cli/tests/e2e_redirect_gem_build.rs and changed only the Driver::ScanVexSourceBlock Gemfile to:

source "<up>/upstream"

source "<up>/upstream" do
  v = if ENV["ZZ"]
    "a"
  else
    "b"
  end
  gem "vuln-gem", "1.0.0"
end

Then I ran:

gem install bundler -v 4.0.22
BUNDLER_VERSION=4.0.22 SOCKET_PATCH_BUNDLER_E2E_VERSION=4.0.22 SOCKET_PATCH_BUNDLER_E2E_REQUIRED=1 \
  cargo test -p socket-patch-cli --test <scratch copy> -- gem_hosted_unwind_keeps_a_source_block_bang --include-ignored --nocapture

Result: rollback exits 0, but the lock differs from the pristine lock only in the missing !:

-  vuln-gem (= 1.0.0)!
+  vuln-gem (= 1.0.0)

The fresh-checkout BUNDLE_FROZEN=true bundle install then fails:

Your lockfile needs to be updated, but it can't be because frozen mode is set.

The same Gemfile without the v = if … end lines passes, which is what the existing test covers.

I also confirmed Bundler's side on its own, with no socket-patch involved. I took a real rubygems.org project (colorize 0.8.1 in the same block shape) and deleted only the !. On Bundler 4.0.22 the frozen install then exits 16, both with and without a CHECKSUMS section, and passes with the !. On 4.0.18 it only warns ("Cannot write a changed lockfile while frozen.") and exits 0.

Expected vs actual

• Expected: CLI_CONTRACT.md / docs/ecosystems.md ("Hosted unwind") say the restore re-derives the upstream lock, and Hosted gem rollback / remove strips the DEPENDENCIES ! of a gem the user declared inside a source "https://rubygems.org" do block, so every frozen install fails after the unwind #1056's fix made the ! survive for gems declared in a user source block. The lock should come back byte-identical (! included), and a frozen install of the restored pair should pass.
• Actual: the ! is dropped and the frozen install fails.

Matrix (Linux, Ruby 3.3.6, main 15c2c83)

macOS and Windows weren't probed. The logic is text-only and OS-independent.

Suspect code

crates/socket-patch-core/src/patch/redirect/upstream/gem.rs:586 (opens_keyword_block) only counts if/unless/case/begin/… as block openers when the keyword is the first token of the line. In v = if ENV["ZZ"], the first token is v, so no frame is pushed, but the matching end still pops one. That pop removes the user's source frame in in_source_block (gem.rs:541), and restore_one falls through to DepEntry::Unpin (gem.rs:899). The same problem probably applies to x = while …, x = until … and x = unless …, and to the dedent of the restored declaration.