fix(gem): bundler vendored+hosted audit — 13 confirmed bugs fixed, tests hardened

[Mikola Lysenko (mikolalysenko)]

Full ULTRACODE review + test matrix over the gem/bundler ecosystem (vendored and hosted modes, all configurations). 24 findings survived adversarial verification; the 13 well-scoped ones are fixed here, each pinned by a test that fails without its fix. Four follow-up PRs (plugin bootstrap, apply/rollback policy, vendor integrity, coverage hardening) will stack on this branch.

Hosted redirect (patch/redirect/mod.rs, gem section)

• Commented-duplicate corruption (P1): the Gemfile edit replaced the first substring occurrence of the matched line, so a commented-out copy above the live line got rewritten, corrupting the Gemfile. Now splices by regex byte range.
• Re-run nesting on grant rotation (P1): the idempotency guard keyed on the exact index URL, but the grant token rotates per request — a re-run nested a second source block. Guard is now grant-agnostic; a rotated grant refreshes the URL in place (new edit kind redirect_gemfile_source_url).
• Attested no-op redirects (P1): gem-level git:/github:/path:/source: options survive into the Socket source block and override it in bundler's DSL — the redirect did nothing yet was confirmed and VEX-attested. Now fails closed with redirect_gem_source_option.
• Platform-suffixed CHECKSUMS siblings now fail closed (redirect_gem_platform_unsupported) instead of inserting a duplicate bare-coordinate pin (P2).
• Paren/tab/multi-space gem declarations are recognized; the append branch is gated on the gem being genuinely undeclared — no more duplicate declarations bundler rejects (P2).
• The lock CHECKSUMS pin is gated on the Gemfile source redirect actually landing (a mixed state guaranteed a checksum failure) (P2).
• New warning that a redirected pair breaks BUNDLE_FROZEN/--deployment installs (pnpm-warning parity) (P2).
• The CHECKSUMS edit now records the upstream sha as original so a future revert can restore it (golden fixture updated). Note: the depscan TS twin (registry-rewrite gem rewriter) may need the same original field for cross-language contract parity (P2).

Vendored backend (vendor/gem.rs, registry_fetch.rs)

• Lock churn with 2+ vendored gems (P1): PATH sections were appended in vendor-invocation order; bundler regenerates them sorted by source identifier. Insert position now matches bundler's sort — verified against a real bundle lock under 4.0.15.
• Re-vendor on a patch update was refused (P1): a new patch uuid for an already-vendored gem hit gemfile_declaration_not_editable — the backend rejected its own wiring, breaking the documented automatic re-vendor. Now rewires in place (direct + transitive shapes).
• Gem lockfile auto-fetch was entirely dead (P1): fetched gems staged into a dir literally named gem, which the platform guard refused as platform_gem_unsupported. Now stages <name>-<version> (with a defense-in-depth coordinate re-check before any I/O).
• Platform-suffixed GEM-specs siblings on pre-2.6 (no-CHECKSUMS) locks now fail closed, mirroring the CHECKSUMS guard (P2).

Crawler + plumbing

• parse_dir_name_version prefers the last dotted-version boundary: http-2-1.0.1 parses as http-2@1.0.1, not the ghost http@2 — real gems (http-2, http-2-next) were never discovered (P1). Known residual: ancient dotted platform suffixes (e.g. sparc-solaris-2.8) mis-split; modern platforms are unaffected.
• vendor/bundle discovery enumerates engine dirs (jruby, truffleruby) instead of hardcoding ruby/ (P2).
• run_nested_apply now threads --ecosystems: scan --ecosystems gem --sync no longer applies (or mutates) other ecosystems' patches (P2).
• scan --vendor --dry-run --vex no longer writes the VEX file nor exits 1 on not-yet-vendored state (P2).

Test hardening

• e2e_gem lifecycle harness: BUNDLE_PATH replaces bundle install --path (removed in bundler 3+); all 3 lifecycle tests green under host bundler 4.0.15.
• The hosted production gem leg now asserts the CHECKSUMS digest actually changed after redirect (client-verifiable today) and, in the success arm, verifies installed bytes against the published afterHashes — an inert gem patch can no longer stay green (the npm minimist blindspot, closed for gem).
• docker_e2e_gem serves the true git-blob beforeHash so the container chain exercises the default non-forced apply path, not just --force.
• setup_matrix_gem module doc corrected: the with-setup Docker cases ARE still a baseline gap (bootstrap deadlock; fix coming in the follow-up plugin PR).

Verification

• socket-patch-core 2460/0, socket-patch-cli lib 350/0, clippy + fmt clean
• e2e_gem 11/11 (incl. live lifecycle under bundler 4.0.15), e2e_vendor_gem_build 6/6 (incl. real-bundler capstone), docker_e2e_gem 2/2, docker_e2e_vendor_gem 1/1 (fresh image)
• hosted production gem leg green: redirect verified against live prod; install still blocked by the known depscan#23630 compact-index 404 (not_built) — server-side, re-verified 2026-08-13

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes touch gem redirect, vendored lock surgery, and nested apply scoping—high user impact for bundler workflows but bounded to gem ecosystem with extensive new regression tests; hosted redirect fail-closed paths may skip redirects users previously thought applied.

Overview
This PR tightens gem/bundler behavior across hosted registry redirects, vendored lock editing, discovery, and CLI plumbing, with tests that fail without each fix.

Hosted Gemfile/Gemfile.lock redirect now splices Gemfile edits by byte range (avoids corrupting commented duplicates), recognizes rotated grant tokens without nesting source blocks, and fails closed on git:/path:/source: options, platform CHECKSUMS siblings, and lock pins without a matching Gemfile redirect. CHECKSUMS edits record upstream sha256 as original for revert; a new warning flags frozen/deployment breakage after redirect.

Vendored gem backend inserts PATH sections in bundler’s sorted order, rewires in place on patch uuid updates (direct and transitive), and stages lockfile auto-fetch as {name}-{version} instead of a generic gem dir. Platform-suffixed GEM specs fail closed on older locks without CHECKSUMS.

Ruby crawler prefers the last dotted version boundary for names like http-2, and discovers vendor/bundle/jruby/truffleruby engine layouts—not only ruby/.

CLI: nested apply after download/scan forwards --ecosystems so scoped sync does not apply other ecosystems; scan --vendor --dry-run --vex skips VEX generation and returns a JSON skipped marker instead of writing files or failing on not-yet-vendored projects.

Tests: BUNDLE_PATH for bundler 4+, hosted gem leg asserts CHECKSUMS change and on-disk afterHashes; docker gem e2e uses real beforeHash without --force; in-process tests cover ecosystem scoping and vendor dry-run VEX.

^{Reviewed by Cursor Bugbot for commit 10c0079. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: CRLF re-run nests Gemfile sources
  • Changed regex pattern from do\n to do\r?\n to match both Unix and Windows line endings, preventing the idempotency guard from failing on CRLF files.

Or push these changes by commenting:

@cursor push 8058d388a7

Preview (8058d388a7)

diff --git a/crates/socket-patch-core/src/patch/redirect/mod.rs b/crates/socket-patch-core/src/patch/redirect/mod.rs
--- a/crates/socket-patch-core/src/patch/redirect/mod.rs
+++ b/crates/socket-patch-core/src/patch/redirect/mod.rs
@@ -1786,7 +1786,7 @@
             let block_re = Regex::new(
                 &(String::from(r#"(?m)^source "("#)
                     + &url_pat
-                    + r#")" do\n  gem ["']"#
+                    + r#")" do\r?\n  gem ["']"#
                     + &regex::escape(&dep.name)
                     + r#"["']"#),
             )

_{You can send follow-ups to the cloud agent here.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 10c0079. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Both Bugbot findings addressed in d063cb1:

• CRLF re-run nests Gemfile sources (High): the block recognizer now accepts do\r?\n, so a core.autocrlf checkout of a previously-redirected Gemfile is recognized instead of re-wrapped. Pinned by gemfile_rerun_on_crlf_checkout_never_nests — verified red without the fix.
• Dry-run VEX skip is JSON-only (Medium): the guard moved to the top of both embed_vex_into_json and embed_vex_human, so every scan arm skips VEX on --dry-run — including the interactive vendor path and the JSON --apply fall-through, which had the same hole. Pinned by two new in-process tests, both verified red without the guard.

Note: took the guard-relocation fix rather than the prepared one-line autofix so the interactive and JSON-apply arms are covered too.

🤖 Generated with Claude Code

[Mikola Lysenko (mikolalysenko)]

Rebased fix/gem-bundler-audit onto origin/main (4e5288e, #164) — new tip 99f9cd0 (was d063cb1 on b109645; main gained 11 commits).

Conflicts resolved (1):

• crates/socket-patch-cli/tests/in_process_scan.rs — purely additive: both sides appended tests at end-of-file. Kept both: main's scan_discovers_maven_and_nuget_in_every_mode (feat: promote maven and nuget to fully available — retire the SOCKET_EXPERIMENTAL_* gates #170 maven/nuget GA) and this branch's two vendor dry-run VEX tests. No behavior judgment calls needed.

Auto-merges verified semantically:

• scan/vendor_flow.rs: main's --vendor-source config threading (fix(scan): honor --vendor-source in the vendored flow #160) and this branch's dry-run VEX skip coexist cleanly (different regions of the JSON path).
• scan/mod.rs: main's takeover warnings (fix(scan): warn when a mode takeover leaves the other ledger stale #164) + hosted JSON envelope (fix(scan): keep hosted --json envelope schema-consistent across discovery #163) coexist with this branch's embed_vex_* dry-run guards and nested --ecosystems threading; all VEX embeds route through the guarded helpers, so every scan arm keeps the dry-run skip.

Verification (all green):

• cargo test -p socket-patch-core full: 2468 passed, 0 failed (redirect goldens included)
• cli --features docker-e2e,setup-e2e: --lib 367 ok; in_process_scan 27 ok (both conflict-side tests run explicitly); in_process_get_manifest_path 5; in_process_get_update_count 3; e2e_gem 8 ok / 3 ignored; docker_e2e_gem 2; docker_e2e_vendor_gem 1; e2e_vendor_gem_build 5 + host capstone (--ignored, ruby 3.4.10 / bundler 4.0.15) 1 ok; in_process_gem_apply 7; in_process_gem_multi_platform 4; scan_vendor_e2e 18; in_process_vendor 27
• cargo clippy --workspace --all-features -- -D warnings (CI-exact): clean
• cargo fmt --check: fails locally, but the failures are pre-existing on origin/main under the pinned toolchain (rustfmt 1.8.0) — the branch's violations are a strict subset of main's (it actually fixes two get.rs ones), i.e. the rebase adds zero new fmt drift. CI has no fmt gate (ci.yml runs clippy only), so left main's files untouched rather than folding out-of-scope reformats into these commits.

🤖 Generated with Claude Code