docs: reframe README to v0.4 (Mac + CUDA)

[FluffyAIcode]

What

Retires the v0.3/v0.2 framing and presents the current, tested engine as v0.4, shipping per platform (v0.4-mac / v0.4-cuda).

• Badge / intro / status / quickstart / architecture diagram → v0.4 (memory-bounded restored Gemma-4 engine; platform tags).
• "What's in the v0.3 architecture" → "v0.4 architecture" — adds restored verifier (Gemma-4 + DFlash + f_θ/S5), per-session binding (PerSessionVerifierRegistry), batched scheduler rows.
• "v0.3 GA evidence" → foundational runtime evidence carried into v0.4 (the 4-h drift / bounded-memory result stays; points to the v0.4 scorecards for the new numbers).
• Single-tenant caveats → v0.4 multi-tenant: PR-A3c done (per-session binding + batched scheduler, 8.45× throughput, recall 1.0).
• MLX section → "v0.4 for Mac"; corrected the stale "spec-decode queued for v0.4" note (it's shipped); CUDA figure updated to 1.79–2.06× AR.
• Roadmap → v0.4 Mac/CUDA/multi-tenant shipped; async continuous batching next.

Historical ADR-index descriptions that legitimately reference v0.2/v0.3 deliverables are left intact (accuracy).

Files

• README.md

Testing

• ✅ Markdown fences balanced; internal anchors point to existing sections.