fix: map GGUF architecture to built-in chat template when embedded is missing

[mvanhorn]

Summary

Maps the GGUF model architecture metadata to a built-in chat template when the model does not ship an embedded template, instead of unconditionally using chatml. gemma/llama/qwen/phi families now get an architecture-appropriate fallback with a specific warning level per case.

Why this matters

Issue #9110 reported that local-inference models without an embedded chat template silently defaulted to chatml, which produces wrong outputs for gemma/llama families. The old llamacpp path issued warn!("Model has no embedded chat template, falling back to chatml") and continued. For a Gemma-family model that is the wrong template; the response degrades badly.

Changes

• New select_chat_template() in local_inference.rs returns either Embedded(t) when the embedded template is present or BuiltIn(BuiltInChatTemplateFallback) when it is not
• Architecture-aware fallback table: gemma/gemma2/gemma3 -> "gemma" (info), gemma4 -> "gemma" with Gemma4Unsupported warning, llama/llama2 -> "llama2", llama3 -> "llama3", qwen2 -> "chatml", phi3 -> "phi3"
• Unknown architecture -> "chatml" with UnknownArchitecture warning
• Missing or whitespace-only architecture -> "chatml" with MissingArchitecture warning
• Architecture matching is case + whitespace insensitive; original-cased name preserved for the log
• LlamaCppBackend::load reads the model arch via meta_val_str("general.architecture") and routes through select_chat_template
• log_chat_template_fallback emits info vs warn based on the fallback variant

Testing

cargo test --features local-inference -p goose --lib local_inference_tests -- 6/6 pass:

• embedded template used unchanged
• 8 known architectures fall back to expected templates with KnownArchitecture warning
• LLaMa3 fallback is case + whitespace insensitive (case preserved in result)
• unknown arch -> chatml + warn
• gemma4 -> gemma + Gemma4Unsupported warn
• missing/empty/whitespace arch -> chatml + MissingArchitecture warn

Fixes #9110

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0d4757cb16

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid treating every llama GGUF as Llama 2

When a GGUF is missing tokenizer.chat_template, this maps the general.architecture value llama to the Llama 2 template, but that metadata value is shared by newer Llama-family GGUFs as well (the local provider’s default model is Llama 3.2). In that fallback path, Llama 3.x models without an embedded template will be prompted with [INST]...[/INST] instead of the Llama 3 header/eot format, leading to malformed prompts and poor or broken responses; the fallback needs more specific model metadata/name handling before choosing llama2.

Useful? React with 👍 / 👎.

[jh-block]

Thanks very much for the work here. For now we have gone a slightly different path (#9442), I wrote a bit about the reasoning in #9110 (comment) — very open to discussion on this as I know the stricter path may make some GGUF conversions (that lack embedded chat templates) a little harder to use, but at least there is an escape hatch for using them now.