output_weight_node_id_ is now always sourced from token_embeddings.weights, which ignores the possibility of an untied LM head (output_weight.weights differing from embeddings). To keep correctness across configs/converted checkpoints, prefer output_weight.weights when present and fall back to token_embeddings.weights only if the output weight file is missing. Also consider reusing the already-mapped embedding node (or mmap_embeddings) to avoid double-mmapping the same file into the graph.

The inline comment in the try block is informal and uses exception handling as control flow. Since “missing projector layernorm” is an expected variant, it would be clearer and cheaper to explicitly check std::filesystem::exists() for those files (and document why) rather than catching all std::exception (which can also mask real corruption/format errors).

Using 0 as a sentinel for “missing” layer-norm weights is unsafe because CactusGraph node IDs start at 0 (the first mmap_weights() call will typically return node_id=0). That means the layer norm can be incorrectly treated as absent even when it loaded successfully. Use a dedicated invalid value (e.g., std::numeric_limits<size_t>::max()), or store presence as a separate boolean/std::optional<size_t> and branch on that instead.

This check uses != 0 to decide whether to apply layernorm, but node_id 0 is a valid graph node ID. If the layer-norm weight happened to be loaded as node 0, normalization will be skipped unintentionally. Switch to an explicit “present” flag or compare against a dedicated invalid sentinel value instead of 0.