2026-04-27 - Phi-4-mini Runtime Scaffolding Smoke Passed

Intent: Continue from completed layer shards and LM-head shards into prompt-ID runtime scaffolding while preserving the ANE-only boundary and validation-before-performance discipline.

Setup: Added converters/phi4_mini_export_runtime.py to export the permitted host embedding lookup bin plus runtime manifest. Added runtime/phi4_mini_ane.swift, a prompt-ID smoke runtime chaining 32 stateful layer shards and 4 ANE LM-head shards, with host work limited to embedding lookup, RoPE/mask bookkeeping, and argmax. Compiled with a local Swift/CoreML compile command.

Result: Wrote local embedding weights (1.1G), the runtime metadata (4.1K), and the runtime binary (148K). The full-chain smoke command loaded all 32 layers plus 4 LM-head shards and generated next token ID 6360.

Surprise / hurdle: Runtime integration could be tested without tokenizer integration by using prompt IDs directly, keeping host work inside the permitted non-compute exceptions.

Lesson: Once all layer and LM-head shards exist, a minimal prompt-ID runtime can prove artifact chaining before spending time on tokenizer, benchmarking, or full-logit golden validation.

Next: No perf/energy benchmarking, tokenizer integration, HF full-logit golden validation, cleanup, or deletion was performed; those remain separate gated steps.

Refs: converters/phi4_mini_export_runtime.py; runtime/phi4_mini_ane.swift; research/ANE_CHAIN_SCHEMA.md