fix: restore local development readiness

[skullcrushercmd]

Summary

• restore a deterministic WSL-first local development workflow for AnyGPT
• pin and bootstrap the ignored LibreChat checkout with its declared Node/npm versions, frozen installs, readiness locks, truthful smoke/doctor modes, and safe process-group cleanup
• repair the Bun dependency graph for Express/ESLint/AJV, restore real wrapper execution, and fix the resulting TypeScript/lint issues without weakening lint policy
• fix the provider error-classification predicate chain and add safe local API/frontend validation paths
• prevent homepage/UI launchers from killing unrelated port owners, with a regression test
• add a tracked, conflict-safe LibreChat customization patch workflow so local React/CSS/asset design changes can be carried through the outer repository

Validation

• bash ./bun.sh install --frozen-lockfile
• bash ./bun.sh run check
• bash ./bun.sh run lint — 0 errors (2,864 existing warnings)
• bash ./bun.sh run build
• bash ./bun.sh run test — API mock/streaming/OpenAI SDK/native coverage plus homepage port-safety regression
• direct LibreChat client tsc --noEmit
• LibreChat full build and both dev and production-style local runtime smoke tests
• keyless/no-provider API + homepage integration smoke
• AnyScan cargo test --locked — 363 passed, 3 live/bench tests ignored
• all 220 package script values parse; all 46 advertised non-service frontend modes pass
• Bash syntax, ShellCheck error severity, git diff --check, staged secret/path/destructive-command scans, and final readiness/port audits pass

Operational notes

• The live production server and its checkout were not accessed, modified, restarted, or deployed during this work.
• The branch was built and tested entirely in the local Debian WSL2 clone.
• The nested LibreChat checkout remains ignored; reusable source/design changes are captured through apps/ui/scripts/capture-customizations.sh and safely replayed by config sync.
• LibreChat's upstream install currently reports 8 audit advisories (6 low, 2 moderate); no unsafe force-upgrade was applied.