Add DMG integrity guard for Linux settings patches

[joshyorko]

Summary

• add a DMG-intel acceptance blocker for unresolved Linux settings patch symbols
• run that integrity scan inside the patcher and carry findings through patch-report.json, so Codex.dmg + patch-report flows see post-patch webview failures
• distinguish real var / let / const declarators from bare assignment expressions in Linux settings icon detection
• update Agent Workspaces settings icon patching and Browser Use node_repl trust detection for current upstream chunk shapes

Verification

• docker run --rm -v "$PWD":/workspaces/codex-desktop-linux -w /workspaces/codex-desktop-linux codex-desktop-linux-devcontainer:local node --test scripts/dev/upstream-dmg-intel.test.js scripts/patch-linux-window-ui.test.js (362/362)
• docker run --rm -v "$PWD":/workspaces/codex-desktop-linux -w /workspaces/codex-desktop-linux codex-desktop-linux-devcontainer:local node --test scripts/patch-linux-window-ui.test.js linux-features/agent-workspace/test.js (378/378, run before the review-fix follow-up; the follow-up did not touch linux-features/agent-workspace)

[avifenesh]

This is an auto review done by revuto.

---

Found two issues with the new DMG integrity guard wiring/detection that could let the intended blocker miss real post-patch failures.

[ilysenko]

Thanks for the PR. I found two blockers in the new integrity guard:

1. The scan is wired to the raw --candidate inventory, so the normal Codex.dmg + patch-report flow cannot see Linux-injected settings symbols from the patched webview assets. It needs to inspect the patched app/extracted ASAR/installed webview assets, or run from the patcher while those files are available.

2. hasLocalPatchSymbolDeclaration() treats bare assignments as declarations, so shapes like ;codexLinuxReadAloudSettingsIcon=e=>... or foo(),codexLinuxReadAloudSettingsIcon=e=>... are skipped even though they have no local declarator. Please distinguish real declarators from expression assignments and add regression coverage for that shape.

[Yo-DDV]

Thanks for the quick follow-up. Small correction from my side: 97a2c01 landed while I was finishing my previous review, so my first note was based on 9a06ba9.

I rechecked the new head. The bare-assignment reproducer now behaves as expected:

• unresolved reference: finding
• real const declaration: no finding
• ;codexLinuxReadAloudSettingsIcon=e=>e: finding
• foo(),codexLinuxReadAloudSettingsIcon=e=>e: finding

Local checks are green on 97a2c01:

• node --test scripts/dev/upstream-dmg-intel.test.js scripts/patch-linux-window-ui.test.js — 362 pass
• bash tests/scripts_smoke.sh
• git diff --check

So the declaration blocker I flagged is fixed in the latest commit. I will leave the broader call to Ilya and CI, but this follow-up looks good for that specific point.

[joshyorko]

@ilysenko addressed both blockers:

• The patcher now scans the patched extracted app while those webview assets exist and writes post-patch integrity findings into patch-report.json; DMG-intel folds those findings into PATCH_INTEGRITY_BROKEN even when --candidate is the raw DMG.
• hasLocalPatchSymbolDeclaration() now only accepts real function/variable declarators. Bare assignments like codexLinuxReadAloudSettingsIcon=e=>... and foo(),codexLinuxReadAloudSettingsIcon=e=>... are covered as failures.

Verified in the devcontainer with node --test scripts/dev/upstream-dmg-intel.test.js scripts/patch-linux-window-ui.test.js passing 362/362.

[ilysenko]

Thanks for the update. I found a blocker:

• This PR touches drift-sensitive patch logic, but it still carries legacy upstream compatibility in the touched paths, including the old agent-workspace Settings visibility fallbacks and older Browser Use node_repl builder handling.
• This repo targets only the latest CODEX.DMG, so please remove those old DMG shapes/fallback paths in the same PR instead of carrying them forward.