GeezerVison

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)

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