Add ADR 0004: Alignment training data preparation policy (Nemotron-informed) (#18)

Locks in the v0.3 alignment training data + LoRA + masking + evaluation
policy before implementation begins. Pure documentation, no code change.

What's new
----------

docs/adr/0004-alignment-training-data-preparation-policy.md

  Decision content (§2):
    2.1 7-domain prompt pool composition (50k prompts)
        chat-en 30%, chat-zh 20%, code 15%, math 10%,
        long-context 10%, multi-turn 10%, tool-calls 5%
        + adversarial/OOD 1k held-out (no hard gate)
    2.2 On-policy verifier rollout config: greedy, sink+window,
        multi-system-prompt rotation, block-aligned hidden state
        capture (block_size=4 deployment)
    2.3 LoRA configuration: o_proj target, rank 128, alpha 512
        (Nemotron-informed) with mandatory A/B/C validation
        before locking in
    2.4 Loss formulation:
          1.0 * smooth_L1 repr alignment
        + 0.5 * KL distill (T=2, top-20)
        + 0.1 * mask recovery
        + position-dependent masking schedule
          p_mask = 0.3 + 0.4 * (pos_in_block / block_size)
    2.5 Per-verifier data isolation policy + versioning convention
    2.6 Greedy-only training assumption (matches ADR 0001 §2.2)
    2.7 Per-slice acceptance gates for v1 ship:
          aggregate >= 0.40 @ K=2
          chat-en >= 0.45, chat-zh >= 0.40, code >= 0.25,
          math >= 0.30, long-context >= 0.30, multi-turn >= 0.35,
          tool-calls >= 0.40
    2.8 Evaluation metrics: acceptance rate, TPF (tokens-per-forward),
        mean acceptance length, speedup vs vanilla AR

  Section 3 — Where Nemotron findings apply (and where they don't):
    3.1 Adopt directly: o_proj-only LoRA, position-dependent masking,
        block-aligned capture, TPF + acceptance length metrics
    3.2 Do not adopt: single-model self-spec architecture, joint AR-
        diffusion pretraining objective, custom CUDA kernels, 14B-scale
        absolute throughput numbers

  Section 4 — Five rejected alternatives with reasons.
  Section 5 — Consequences (positive + accepted trade-offs +
              implications for code).
  Section 6 — Validation criteria for ADR completion.

Sources informing the ADR
-------------------------
  Fu et al. 2026, 'Nemotron-Labs-Diffusion: A Tri-Mode Language Model
  Unifying Autoregressive, Diffusion, and Self-Speculation Decoding'
  (NVIDIA technical report, May 2026).

  arXiv:2512.14067, 'Efficient-DLM: From Autoregressive to Diffusion
  Language Models, and Beyond in Speed' (Fu et al., Dec 2025).

  HuggingFace model card linear_spec_lora subfolder of
  nvidia/Nemotron-Labs-Diffusion-14B (LoRA target/rank/alpha confirmed).

Why ADR 0004 ships now (not when v0.3 starts)
---------------------------------------------
The data-prep choices materially constrain the trainer implementation.
Writing the ADR after starting the trainer means either retrofitting
decisions (waste) or churning the ADR mid-implementation (worse).
Cost of writing this ADR pre-v0.3: ~3 hours of writing. Benefit: cleaner
v0.3 PR sequence with locked-in inputs/outputs.

Documentation updates
---------------------
- docs/adr/README.md: ADR 0004 added to the index.
- README.md: ADR badge updated to '0001 | 0002 | 0003 | 0004';
  ADR list at the bottom gets a one-paragraph summary entry for 0004.

No code change; full test suite still passes (451 passed, 100% coverage).

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: FluffyAIcode

README.md

@@ -3,7 +3,7 @@
 [![CI](https://github.com/FluffyAIcode/Kakeya-LLM-Inference-engine/actions/workflows/ci.yaml/badge.svg?branch=main)](https://github.com/FluffyAIcode/Kakeya-LLM-Inference-engine/actions/workflows/ci.yaml)
 [![Release](https://img.shields.io/badge/release-v0.1.0-blue)](https://github.com/FluffyAIcode/Kakeya-LLM-Inference-engine/releases/tag/v0.1.0)
 [![Platform](https://img.shields.io/badge/platform-Apple%20Silicon-lightgrey)](docs/local-inference-engine.md)
-[![ADRs](https://img.shields.io/badge/ADRs-0001%20%7C%200002%20%7C%200003%20%7C%200006-green)](docs/adr/)
+[![ADRs](https://img.shields.io/badge/ADRs-0001%20%7C%200002%20%7C%200003%20%7C%200004%20%7C%200006-green)](docs/adr/)
 
 Runs the speculative-decoding architecture designed in the prior product
 discussion using **real, public** weights:
@@ -468,6 +468,15 @@ explicitly rejected.
   and what intermediate step ships in v0.2 — `PooledVerifier`
   wrapper that makes pool memory accounting accurate without
   touching the model forward.
+- [ADR 0004 — Alignment training data preparation policy
+  (Nemotron-informed)](docs/adr/0004-alignment-training-data-preparation-policy.md):
+  the v0.3 alignment training data + LoRA + masking + per-slice
+  evaluation policy, locked in before implementation begins.
+  Adopts NVIDIA Nemotron-Labs-Diffusion's `o_proj`-only LoRA
+  configuration (rank 128, α 512) pending an A/B/C validation
+  experiment, plus 7-domain prompt pool, block-aligned hidden
+  state capture, position-dependent masking, and per-slice
+  acceptance gates.
 - [ADR 0006 — Project positioning as local agent
   infrastructure](docs/adr/0006-local-agent-infrastructure-positioning.md):
   the strategic positioning decision that Kakeya is **local agent