docs: record ADR 0004 Phase 1 outcome (export-first validated)

[inureyes]

Records the Track B Phase 0/1 outcome (issue #449, branch spike/openxla-export-449) into ADR 0004, and refines the open items per the 2026-06-27 decisions.

What this records

• Export-first validated for the fp16 path. Llama-3.2-1B authored once as a JAX reference exports to StableHLO via jax.export and greedy-decodes token-exact (48/48) vs HF temp-0; the same module runs on PJRT and IREE with no custom_call. Logits agree with HF to fp32 noise (~1e-5).
• Session-contract shape confirmed. The exported prefill/decode_step graphs with session-owned KV and on-device argmax match the InferenceSession contract landed in refactor: redraw ComputeBackend as an inference-session engine and route the CLI through it #450 (refactor: redraw ComputeBackend as an inference-session engine contract and move the MLX path behind it (byte-identical) #448). Phase 3 binds to it.
• Authoring frontend still open, evaluated in parallel. JAX-reference is proven; a Rust-native StableHLO emitter is under parallel evaluation (feat: evaluate a Rust-native StableHLO emitter as the compiler-family authoring path (spike) #451) to keep compiler-family model definitions in-tree in Rust with no Python in the authoring path. The StableHLO artifact is the swappable boundary, so this does not affect the runtime or the refactor: redraw ComputeBackend as an inference-session engine contract and move the MLX path behind it (byte-identical) #448 integration. torch.export-from-HF remains unexercised.
• GPU deferred; Phase 2 split. The GB10 (aarch64 + Blackwell) box has no PyPI CUDA jaxlib, so Phase 1 ran on CPU. Phase 2a (int4 dequant-in-graph lowering structure + fusion + correctness, frontend-independent) proceeds now on CPU; Phase 2b (GPU perf/memory) is deferred until a GPU host is chosen.
• Per-model cost clarified. 'Per-model shrinks to glue' holds within an architecture family (weights are graph inputs), is scaffold-plus-deltas across transformer variants, and needs new authoring for new families (MoE, SSM-hybrid, VLM).

ADR stays Proposed; it moves to Accepted once the OpenXLA backend integrates behind the #448 session contract (Phase 3). Docs only. References #449, #450, #451, #448, #338.