[RFC 33] Address /rfc-review findings: citations, EDEN attribution, implementation appendix

Citation fixes (BLOCKERs):
- Split ref [14] into three. The arxiv:2604.18555 paper is "A Note on
  TurboQuant and the Earlier DRIVE/EDEN Line of Work" (Ben-Basat et al.,
  April 2026), not the EDEN paper. EDEN is arxiv:2108.08842 (ICML 2022);
  DRIVE is arxiv:2105.08339 (NeurIPS 2021). Re-thread inline citations:
  Note paper for "TurboQuant is suboptimal" claims, EDEN for the
  algorithm and optimal-S criterion, DRIVE as the 1-bit predecessor.
- Mark 0xSero/turboquant (GPL-3.0) as for-illustration-only with explicit
  license note; not a code dependency under Vortex's MIT/Apache rule.
- Replace develop-ref [current-impl] URL with commit-pinned SHA at PR
  #7829 merge (ff120401).
- Pin arXiv versions on TurboQuant (v1), PDX (v1), HIGGS (v1) where the
  RFC quotes from them.

Content fixes (MAJORs):
- Add §4 "Naming" subsection explaining why TurboQuant branding stays
  despite EDEN priority.
- Specify EDEN-S adoption at implementer level: cite EDEN [15] for the
  optimization criterion + reference implementation
  (amitport/EDEN-Distributed-Mean-Estimation, MIT).
- Rewrite §10 "Integration with Vortex" with the real symbols from
  vortex/examples/turboquant_vector_search.rs, including both the
  compressor-builder opt-in path and the explicit-encode path.
- Add Appendix D "Implementation Specification" consolidating the
  wire-format invariants, prost metadata schema, encode/decode
  pseudocode (Stages 1-2), PDX kernel, error model, worked example,
  test plan, performance budgets, registry wiring, crate boundaries,
  and migration sequence.
- Resolve four "RESOLVABLE NOW" Open Questions items (norm storage
  shape, EDEN-S adoption timing, EDEN-vs-QJL unbiased path,
  vortex-tensor cleanup) — document the resolutions in §6/§7/§15/§14.

Minor and nit fixes:
- Reconcile SORF FLOP count methodology with the original RFC's table
  (drop normalization multiplies from the count, per the original).
- Restore PDX implementation file-path citations
  (include/pdx/quantizers/scalar.hpp etc.) under reference [4].
- Disambiguate "chunk" (PDX 64-vector group) from "block" (TurboQuant
  block_size coordinates) in §8.
- Add a Drawbacks section consolidating non-unit-norm decode, L2 norm
  O(1)→O(k) regression at Stage 2 k>1, dual-implementation maintenance
  during migration, and other accepted costs.
- Add an architecture diagram to §5 showing the relationship between
  Vector input/output, the extension type, derived state, and storage.
- Move pseudocode out of §7/§8 into Appendix D.

Signed-off-by: Will Manning <will@willmanning.io>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Will Manning <will@willmanning.io>

Author: lwwmanning