Fix gem source-block bang lost after x = if … end (#1476)

[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1476

Summary

Hosted gem rollback / remove now keep the DEPENDENCIES ! (and the
declaration's indent) for a gem declared in the user's own
source "…" do block when an assignment-form keyword construct
(v = if … end, v = case … end, v = begin … end, x ||= unless …)
precedes it. Before this, the restored lock dropped the ! and every
frozen/deployment install on Bundler 4.0.19+ failed with exit 16, while the
unwind reported success.

Root cause

in_source_block (crates/socket-patch-core/src/patch/redirect/upstream/gem.rs)
decides DepEntry::Keep vs Unpin by counting block openers and ends
line by line. opens_keyword_block only counted if/case/begin/…
when the keyword was the first token of the line. In v = if … nothing was
pushed, but the matching end still popped, removing the user's source
frame. declaration_prefix used the same line-level heuristics, so the
restored gem also came back at column 0 after a blank line.

Fix

One block scanner (block_stack) now feeds both in_source_block and
declaration_prefix:

• word-level: a keyword opens a construct wherever it starts an expression
  (line start, after =, (, ,, ;, a label require:, =>, a
  ternary, and/or/when/then/else, an operator, or a
  continued previous line); end closes one; one-line constructs net out;
• modifiers (gem "x" if c, x.save! if y, return if y, a modifier on a
  \\-continued line), names (x.class, :if, if:, begin?,
  def begin) and endless defs (def x = 1, def ==(o) = …) open nothing;
  a loop's own do is not a second block;
• strings, backticks, %w[]/%i()/%q{}/%x()/… and regexp literals
  (multi-line strings and %-literals included), comments, =begin/=end
  blocks and heredoc bodies are not code;
• the restored declaration's indent comes from the enclosing block's opener
  line (+2); a blank line is still restored before a following declaration.

Tests (red → green)

Issue case	Test	Before fix	After
#1476 v = if … end lock ! + Gemfile byte-identical	unit source_block_declaration_keeps_its_bang_through_restore (new manifests: if, case+begin, unless/while, deeper end)	FAILED (\n\ngem "rails" at column 0, ! dropped)	ok
#1476 scanner shapes (if/unless/case/begin/while/until/for, labels, ternary, literals, heredocs, =begin)	unit source_block_detection	FAILED	ok
layout after an inner end	unit manifest_blocks_restore_in_place	—	ok
#1476 e2e on real Bundler (v = if … end + w = case … end), rollback + remove, frozen install of the restored pair	e2e_redirect_gem_build::gem_hosted_unwind_keeps_a_source_block_bang (new ScanVexSourceBlockAssignment driver)	issue repro: exit 16 on 4.0.22	ok on Bundler 4.0.22 and 4.0.18

Commands run locally (Ruby 3.3.6):

• cargo fmt --all -- --check ✅
• cargo clippy --workspace --all-features -- -D warnings ✅
• cargo test -p socket-patch-core --all-features --lib redirect::upstream::gem: 17 passed; with the fix reverted, 2 failed
• BUNDLER_VERSION={4.0.22,4.0.18} SOCKET_PATCH_BUNDLER_E2E_REQUIRED=1 cargo test -p socket-patch-cli --all-features --test e2e_redirect_gem_build --test e2e_vendor_gem_build -- --ignored: 29 + 8 passed on each
• cargo test --workspace --all-features --no-fail-fast: 14337 passed, 13 failed. All 13 are environmental and outside gem code: 12 use chmod to force write failures, which root ignores (the sandbox runs as uid 0), and e2e_bun_lockb::binary_shared_bundled_record_hosted_pin_is_managed needs the real patch API (tunnel error). CI runs these unprivileged.

No wrapper (npm/, pypi/, gem/) changes: this is core Rust logic only.

Review notes (/code-review high, two passes)

All confirmed findings are fixed in c9f3945 and 9f2c94f. Not changed:

• vex/discover/gem.rs parse_source_block_decls keeps its own line
  grammar. It reads only Socket-written patch-registry blocks (which hold
  plain gem lines) and fails closed: an early close only drops
  attribution and never adds a ! or an attestation. Sharing the lexer
  there would be a separate refactor.
• block_stack runs twice per restored direct gem (in_source_block and
  declaration_prefix) over a small Gemfile prefix. Threading one stack
  through restore_one → restore_manifest costs more code than it saves.
• ruby_words looks like gem_line_tail_blocks_edit's classifier in
  redirect/mod.rs, but the two answer different questions (modifier vs
  argument position on a declaration line), so they stay separate.
• Remaining text-lexing limits, same as before this PR: #{} interpolation
  that nests quotes, and percent literals after a bare method call
  (puts %w[…]).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HuKqLAEGGwH73DucozCrY6

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

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

^{Reviewed by Cursor Bugbot for commit 9f2c94f. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Burn-down agent: labeled Ready for review at head 9f2c94f.

• CI: 35/35 green (+50 skipped) on 9f2c94f; no main-wide failures.
• Bugbot: reviewed 9f2c94f, no findings (a later re-request was cancelled, not failed).
• Mergeable, no conflicts; does not touch CHANGELOG.md.
• Reviewer focus: crates/socket-patch-core/src/patch/redirect/upstream/gem.rs keyword/literal scanning for assignment-form x = if … end inside source blocks.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief (merge steward)

What it does: Hosted gem rollback/remove decides whether a restored gem sits inside the user's source "…" do block (and so keeps its ! in the lock's DEPENDENCIES) by counting block openers and ends. Before this PR, assignment-form constructs like v = if … end didn't count as openers but their end still closed the source block, so the ! was dropped and frozen Bundler 4.0.19+ installs failed with exit 16. A single word-level Ruby block scanner (block_stack) now feeds both in_source_block and the restored line's indent.

Risk: medium. The change is only in the gem unwind path, but it replaces a small line-level heuristic with roughly 450 lines of hand-written Ruby lexing (literals, heredocs, =begin, endless defs, modifiers vs. expression-position keywords). Unusual Gemfile syntax could still be misread, as it could before. When that happens, only the !/indent of the restored line is affected.

Look here:

• block_stack gem.rs:565:`` the core scanner (do/keyword push, `end` pop, `;` statements, carried continuation lines)
• starts_expression gem.rs:709:``` decides between a modifier (gem "x" if c) and an opener (`v = if c`)
• masked_ruby_code gem.rs:812 and heredoc_openers gem.rs:940:````` literal, comment and heredoc masking
• declaration_prefix gem.rs:512:`````` restored-line indent now comes from the enclosing block's opener. Top-level layout is unchanged.
• e2e gem_hosted_unwind_keeps_a_source_block_bang:``````` real-Bundler rollback/remove plus a frozen install with the new ScanVexSourceBlockAssignment Gemfile

Verified:

• Read the full diff (2 files, +637/−64): scope is limited to gem.rs and its e2e test. No CHANGELOG, wrapper or debug leftovers. CRLF is handled (split_inclusive('\n') + \r trim).
• Checked that the top-level declaration_prefix output (blank line, next declaration's indent) matches the old behavior, so only in-block cases change. Those now get better indents, including multi-line group openers and if blocks.
• Locally on 9f2c94f: cargo test -p socket-patch-core --all-features --lib redirect::upstream::gem gave 17/17 pass. CI's cargo clippy -p socket-patch-core --all-features -D warnings is clean, and so is cargo fmt --check. The --all-targets lints are pre-existing in other files and none are in gem.rs.
• CI: 35/35 green, 50 path-skipped. ci-ok and clippy are green. Bugbot reviewed 9f2c94f with no findings. No unresolved threads.

Changes I made: none.

Open questions: none blocking. Known lexing limits are listed in the PR description: nested-quote #{} interpolation and puts %w[…] after a bare call. Both are the same as before this PR.

Auto-merge (squash) is armed, so approving sends this straight to the merge queue.

---

Generated by Claude Code