Improve atlantis-multilang

[0xdkay]

Problem

• multilang currently only support old AIxCC base builder (v1.3.0) and old models, some of which were deprecated. (e.g., sonnet-3.7).
• Latest oss-fuzz uses clang 22, which is not supported by multilang yet.

Plan

1. make it work on both clang 18 (AIxCC base builder) and clang 22
2. make it support latest models <-- this will also require update on internal dependencies.
3. hopefully refactor large modules init
4. make each module selectable by env vars (additional_env) from compose.yaml

[0xdkay]

Progress Update

Items 2 and 4 from the plan are addressed in atlantis-multilang-wo-concolic#1 (merged) and the registry/example update in #146:

Done

• Item 2: Latest models — Updated to gpt-5.4/5.4-mini, claude-sonnet-4-6/opus-4-6/haiku-4-5, gemini-3.1-pro-preview. Upgraded internal dependencies: langchain 0.3.x → 1.x, all cross-file pip conflicts resolved, Joern v4.0.507, jedi-language-server 0.46.0, latest litellm pricing JSON.
• Item 4: Module selection via env vars — CRS_INPUT_GENS in additional_env controls which modules run (given_fuzzer, mlla, testlang_input_gen, dict_input_gen). Services (redis, joern, lsp) are conditionally resolved based on active input generators.

Additional fixes

• Harness discovery: switched from allowlist to blocklist for fuzz target extensions (fixes @NO_OOM, .vis, .udp harness names)
• get_harnesses() always scans /out/ for binaries, merges config.yaml source mappings (fixes JVM targets like POI where create-config missed harnesses)
• Added c++, jvm to supported_target.language and undefined, memory to sanitizers
• Fixed StructuredTool import, missing dataclasses import for langchain 1.x

Remaining (future work)

• Item 1: Clang 22 / latest oss-fuzz base image — Still on AIxCC base-builder v1.3.0 (clang 14/18). DynamoRIO 11.91 upgrade was attempted but requires GLIBC 2.34 (base image has 2.31).
• Item 3: Refactor large modules init — Dead code removed (constraint-gen, concolic/symcc), but no structural refactoring yet.