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

Summary

formats::gem::manifest::classify decides whether BUNDLE_GEMFILE names the project's own Gemfile / gems.rb by comparing std::path::absolute + normalize_lexically of the setting against the project root. It never resolves symlinks. The root comes from the process cwd (getcwd, the physical path) or --cwd. A shell's $PWD, and paths people type, are often the logical path through a symlink. So BUNDLE_GEMFILE=$PWD/Gemfile names the same file Bundler loads, but socket-patch classifies it as "another manifest" (LoadedManifest::Unsupported).

On macOS this is the default for anything under /tmp (→ /private/tmp) or $TMPDIR (/var/folders/… → /private/var/…). On Linux it hits any project reached through a symlinked directory (a symlinked workspace or home, /app → volume, and so on).

Impact

All of these fail closed, but each one stops socket-patch from working on a correctly configured project, with a message that's factually wrong:

• Hosted scan redirects nothing: redirect_gem_bundle_gemfile_unsupported ("bundler loads /tmp/bh-app/Gemfile … not the project's Gemfile or gems.rb"). That is the project's Gemfile.
• vex on an already-redirected project whose install is patched (Bundler loads the patched gem) refuses with exit 2: "BUNDLE_GEMFILE points bundler at another manifest, so this wiring is never installed and the patch is not attested". It also suggests deleting Gemfile.lock, which is the lock Bundler uses.
• rollback / remove error out with patched_ref_unattributable for the same reason, so the hosted patch can't be unwound while the variable is set.

The same function backs bundler_loaded_lock_in (lock inventory, ledger recovery, VEX discovery) and the vendored manifest check, so every gem lock reader inherits this.

Repro (Linux; the ln -s stands in for macOS /tmp)

mkdir -p real/app && ln -s "$PWD/real" link && cd link/app
printf 'source "https://rubygems.org"\n\ngem "colorize", "0.8.1"\ngem "rainbow"\n' > Gemfile
bundle config set --local path vendor/bundle && bundle install && bundle lock --add-checksums
echo "PWD=$PWD physical=$(pwd -P)"

BUNDLE_GEMFILE=$PWD/Gemfile bundle exec ruby -e 'puts Bundler.default_lockfile'   # → link/app/Gemfile.lock (same file)

A="--api-url <mock> --org org --api-token fake --patch-server-url <mock>"
socket-patch scan --mode hosted --json --yes --dry-run $A                              # redirected: 1
BUNDLE_GEMFILE=$PWD/Gemfile socket-patch scan --mode hosted --json --yes --dry-run $A  # redirected: 0, redirect_gem_bundle_gemfile_unsupported
BUNDLE_GEMFILE=$(pwd -P)/Gemfile socket-patch scan --mode hosted --json --yes --dry-run $A  # redirected: 1

# after a real redirect (no env) + bundle install → installed gem is patched:
BUNDLE_GEMFILE=$PWD/Gemfile socket-patch vex --product pkg:gem/app@1 $A      # exit 2 ("never installed")
BUNDLE_GEMFILE=$PWD/Gemfile socket-patch rollback --dry-run --json --yes $A  # status error, patched_ref_unattributable

The same refusal happens for --cwd <symlinked path> with BUNDLE_GEMFILE=<physical path>, and for .bundle/config BUNDLE_GEMFILE: "<symlinked abs path>/Gemfile" (as written by bundle config set --local gemfile "$PWD/Gemfile"), with no environment variable at all.

Expected vs actual

• Expected, per docs/ecosystems.md (RubyGems row): BUNDLE_GEMFILE "is followed when it names the project's Gemfile / gems.rb, and any other configured manifest is refused". Bundler expands the path and resolves it to the same file and the same Gemfile.lock (Bundler.default_lockfile above), so socket-patch should treat it as the project's Gemfile.
• Actual: the refusal fires when it shouldn't (a false Unsupported), and the VEX / rollback messages claim Bundler uses a different manifest.

Matrix (Bundler 4.0.22)

Probe run: https://github.com/SocketDev/socket-patch/actions/runs/37382646373 (the probe's own vex/rollback cells hit a stale vendor/bundle harness artifact; the Linux sandbox rows cover them).

First bad commit

9d718cf5 (#431, the fix for #341 / #390), which introduced the BUNDLE_GEMFILE classification with a lexical compare; cbf1f748 (#532) kept it. The v4.0.0 release binary ignores BUNDLE_GEMFILE and redirects this project (it predates the classification), so this hasn't shipped in a release yet.

Suspect code

• crates/socket-patch-core/src/formats/gem/manifest.rs:138-147 (resolve_against: absolute + normalize_lexically only) and :183-189 (target == root.join(manifest)).
• The same lexical compare is in env_keeps_root (manifest.rs:163-166), which decides whether the env var moves Bundler's root.

Comparing canonicalized paths (or file identity, same_file-style) whenever both exist would match what Bundler does. A lexical match could stay as the fast path, falling back to canonicalization only when it fails, so a symlinked Gemfile file (already refused elsewhere) keeps its own handling.

Backlog review — 2026-10-08

Priority: P1 → P2. Lexically different spellings of the same Gemfile cause a refusal; canonical path spelling works.