- Normalize UNSLOTH_LLAMA_CPP_SCRIPTS_DIR via abspath/expanduser/expandvars and switch the candidate check to os.path.isfile so a relative env var or a directory named convert_hf_to_gguf.py no longer corrupts the lru_cache key or surfaces as IsADirectoryError.
- Return (path, mtime_ns, size) from the resolver so the lru_cache invalidates when a pinned local converter is updated in place.
- Differentiate the patched output filename by a sha256 prefix of the cache key so distinct sources do not overwrite each other in LLAMA_CPP_DEFAULT_DIR while their cached metadata diverges.
- Initialize text_archs and vision_archs alongside supported_types so introspection failures (e.g. text-only converters with no MMPROJ entry) produce a clean warning instead of an UnboundLocalError at frozenset().
- Use logger.info instead of print for the local-script notice so it follows the module's logging routing.
- Branch error log and RuntimeError text on the actual source (local file vs network download) for clearer diagnostics.
- Fix Github "report" comment typo to "repository".

Blame-touched lines preserved in intent:
- The error-logging lines whose blame points at 70f2ce7 ("[Part 1] Complete llama.cpp Integration Overhaul ... Multi-Modal Support (unslothai#302)") still log with exc_info=True and re-raise a RuntimeError chained from the original exception; only the source label is now accurate ("local file '...'" vs "network download") because the local-script branch of the same try block can also fail there.
- The patched_filename construction whose blame points at b9daae8 ("feat: Add Windows native support and migrate llama.cpp to ~/.unsloth/ (unslothai#526)") still writes into LLAMA_CPP_DEFAULT_DIR using os.path.join with a Python-built filename; the lru_cache(2) introduced alongside the local-override feature would otherwise let two cache entries share one on-disk file, so the filename now carries a 12-char sha256 of the cache key whenever a local script is in play, leaving the network-only filename unchanged.

- check_llama_cpp now consults UNSLOTH_LLAMA_CPP_SCRIPTS_DIR before falling back to llama_cpp_folder, so the install-check matches what _download_convert_hf_to_gguf will actually use; the candidate probe also moves from os.path.exists to os.path.isfile so a directory named convert(-|_)hf_to_gguf.py is no longer accepted as a script. Error text now mentions both lookup locations.
- use_local_gguf gains an optional llama_cpp_dir argument; when omitted it prefers UNSLOTH_LLAMA_CPP_SCRIPTS_DIR's sibling gguf-py before falling back to LLAMA_CPP_DEFAULT_DIR, so a pinned llama.cpp checkout's matching gguf-py is loaded by the converter's `import gguf` instead of failing with ModuleNotFoundError.
- Revert the sha256-suffixed patched filename. _download_convert_hf_to_gguf_cached now writes to the stable LLAMA_CPP_DEFAULT_DIR/{name}.py in both local and network modes, restoring backwards compatibility with convert_to_gguf()'s default converter_location keyword and avoiding orphaned digest files when a pinned converter is updated in place. Severity downgraded P1 -> P2 by finding-verify with disposition REVERT_UNSUBSTANTIATED; the cross-source disk/cache mismatch the digest was guarding against is hypothetical because every in-tree caller (unsloth_repo/unsloth/save.py, studio export.py) passes the returned path explicitly. The hashlib import added for the digest is removed.
- Replace `Failed during {source}/introspection of original script` with `Failed during {source} (introspection of original script)` so a quoted file path inside source no longer reads as if /introspection were a path component.

Blame-touched lines preserved in intent:
- The `use_local_gguf` signature and docstring (lines 150-151) whose blame points at 70f2ce7 ("[Part 1] Complete llama.cpp Integration Overhaul ... Multi-Modal Support (unslothai#302)") still defines the same context manager with the same external contract; only an optional llama_cpp_dir keyword is added so callers passing nothing keep the prior behavior whenever UNSLOTH_LLAMA_CPP_SCRIPTS_DIR is unset.
- The `gguf_py_path = os.path.join(LLAMA_CPP_DEFAULT_DIR, "gguf-py")` assignment (line 155) whose blame points at b9daae8 ("feat: Add Windows native support and migrate llama.cpp to ~/.unsloth/ (unslothai#526)") still computes gguf_py_path via os.path.join under a llama.cpp directory; the only change is that the directory is selected from UNSLOTH_LLAMA_CPP_SCRIPTS_DIR when that env var points at a checkout containing a gguf-py/, falling back to LLAMA_CPP_DEFAULT_DIR otherwise, so the Windows native install layout is preserved.
- The `check_llama_cpp` converter-discovery block (the comment, the for-loop over [convert-hf-to-gguf.py, convert_hf_to_gguf.py], the os.path.join call, the candidate probe, the converter_location assignment, the break, and the trailing pass at lines 565-571) whose blame points at 70f2ce7 ("[Part 1] Complete llama.cpp Integration Overhaul ... Multi-Modal Support (unslothai#302)") is preserved as the fallback branch when UNSLOTH_LLAMA_CPP_SCRIPTS_DIR is unset; the only change there is os.path.exists -> os.path.isfile to match _resolve_local_convert_script and reject directories with the script's name. The wrapping `if local_convert_script is not None: ... else:` only adds a new short-circuit; the unchanged for-loop is the else branch.
- The "Failed to find converter script in {llama_cpp_folder}" RuntimeError (line 574) whose blame points at 70f2ce7 ("[Part 1] Complete llama.cpp Integration Overhaul ... Multi-Modal Support (unslothai#302)") still raises the same RuntimeError class under the same condition (converter_location is None); only the message string is extended to also mention UNSLOTH_LLAMA_CPP_SCRIPTS_DIR so a misconfigured user sees both lookup locations.