Upstream contribution from apr-cookbook architecture-demos campaign. Lifts the cookbook's PMAT-309 cross-family detector primitive into aprender-core as a public API.

What's added

// Before: cookbook had its own discriminator dispatcher
let v = inference_arch_detector::detect("config.json");

// After: cookbook calls upstream
use aprender::format::FamilyRegistry;
let family = FamilyRegistry::detect_from_config_str(&body);

detect_from_config_str(body: &str) -> Option<&'static str>

Discriminator-field-based dispatch from raw config.json body to family name. Order-sensitive priority list:

• Qwen3.5 (tie_word_embeddings + head_dim + qwen3_5) → before Qwen3
• Qwen3 (head_dim + qwen3, NOT qwen3_5) → before Qwen2
• Qwen2 (qwen2 + rope_theta)
• Phi (qkv_proj_fused)
• Gemma (query_pre_attn_scalar)
• GPT-NeoX (use_parallel_residual)
• OPT (do_layer_norm_before)
• GPT-2 ("n_embd")
• OpenELM (ffn_multipliers + num_query_heads)
• DeepSeek (n_routed_experts)
• Falcon-H1 (mamba_d_state + mamba_expand + falcon_h1)
• RWKV-7 (time_mix_extra_dim)
• MAMBA (state_size + conv_kernel, NOT num_attention_heads)
• BERT (type_vocab_size)
• Mistral (sliding_window + MistralForCausalLM) → before Llama
• Whisper (WhisperForConditionalGeneration)
• Moonshine (MoonshineForConditionalGeneration)
• Llama (LlamaForCausalLM) — catch-all, checked LAST

Alias mechanism (unblocks 11+ blocked families)

let mut registry = build_default_registry();
registry.register_alias("codellama/*", "llama")?;
registry.register_alias("TinyLlama/*", "llama")?;
registry.register_alias("HuggingFaceTB/SmolLM-*", "llama")?;
// ...

let family = registry.resolve_alias("codellama/CodeLlama-7b-hf");
// Some("llama")

Unblocks the 11 alias entries from apr-cookbook's architecture-demos/manifest.yaml: codellama, dolphin, hermes, openchat, smollm, smollm2, tinyllama, vicuna, wizardcoder, yi, zephyr — all derive from llama or mistral and just needed pattern aliasing.

Tests

• 9 detector_tests — one per supported family + edge cases (priority order, unknown configs, determinism)
• 5 alias_tests — glob matching, exact matching, parent-validation, count, resolution

All 14 pass; no regressions in existing aprender-core suite.

Why upstream

apr-cookbook's PMAT-309 inference_arch_detector recipe was effectively a reverse-engineered discriminator catalog living in cookbook code. Lifting it here:

1. Gives the catalog an authoritative home (the contracts/model-families/ upstream)
2. Lets the cookbook recipe become a thin demo of aprender::FamilyRegistry::detect_from_config_str instead of a parallel implementation
3. Provides the alias mechanism the cookbook tracked as upstream backlog (25 status: blocked entries in the manifest)

Follow-up PR will update apr-cookbook to consume this API.

🤖 Generated with Claude Code