[bajrangCoder]

Fixed

• Command helpers are now followed when the Executor sink sits in a callback inside them, e.g. function exec(cmd) { return new Promise(r => Executor.start(cmd, ...)) }. Before this, every command passed to such a helper was lost, including curl ... | sh, and only a generic "built at runtime" finding was reported.
• btoa(...) of a known string, and local base64 helpers that return btoa(...), now fold to their value.

Added

• shell.encoded_script (Medium): base64 piped through base64 -d inside a command (printf %s <b64> | base64 -d > file, echo <b64> | base64 -d | sh) is decoded, and the decoded script gets the same checks as any command.
• shell.staged_script (High): a file written with content the scanner can't read, then run with sh <file> from Executor or a terminal tab. The file's multi-line, script-shaped constants are checked as the likely content.
• shell.builds_remote_source: git clone followed by a build step (make, cmake, npm install, cargo build, …), pip install git+…, and npm install github:….
  • The evidence lists each repository, and expands for R in a b c; do git clone …/$R fallback lists.
  • Medium when unpinned; Low when pinned (--branch <tag>, git checkout <sha>, pkg@<ref>).

Other

• RULES_VERSION is bumped to 2026.11.4.
• The README covers the new behaviour and rule IDs.

Tests

• cargo test: 88 pass. New tests cover:
  • a Promise-wrapped helper
  • btoa and base64-helper folding
  • inline base64 decoding
  • hidden staged scripts, and the case where a written file is never run
  • pinned and unpinned source builds, including package-install forms
  • script-shaped constant detection
• cargo clippy --all-targets is clean.
• Before/after scans of three real plugin zips: two are unchanged; the third gains the expected staged_script and builds_remote_source findings, and its helper call sites are graded individually.