fix(whitepaper): v1.1 - Correct DLFloat attribution and remove overclaims

Critical corrections based on literature review:

1. GF16 = IBM DLFloat (1/6/9 format, Agrawal 2019)
   - Add honest citation: format is NOT novel
   - Our novelty: integer-backed u16 implementation + FPGA characterization

2. Remove φ-optimality claims
   - IBM arrived at 1/6/9 via distribution analysis, not golden ratio
   - Keep format as historical motivation only

3. Fix energy calculation
   - Remove incorrect "10× energy savings" claim
   - Correct to 2× model-based estimate (aligned with literature)
   - Add caveat: actual hardware measurements required

4. Add proper references
   - Agrawal 2019 (DLFloat original)
   - Mellempudi 2021 (formal proof of 1/6/9 optimality)
   - Schröder 2024, Van den Berg 2024 (FP8/energy)
   - Zhou 2023, TerEffic 2025 (ternary FPGA)

5. Clarify novelty claims (what we do vs don't claim)
   - DO claim: u16 implementation, FPGA characterization, ternary comparison
   - DON'T claim: format novelty, φ-optimality, 10× energy

This makes the whitepaper scientifically honest and peer-review-ready.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Antigravity Agent