v0.11.2 memory-plus axes 5/6 + SBPE — first axis to beat plain zlib

Adds four new aged-body compaction paths on top of v0.10.0 axes 1-4:

- Axis 5 (OMCD): explicit-base delta compression. Stores text as base_hash +
  prefix_len + suffix when the new text shares ≥64 bytes of prefix with an
  existing entry. Verified 1899→65 bytes on pure-append. New MCP tool
  omc_memory_store_delta.

- Axis 6 (OMCH): HBit dual-band split — substrate-tokenize + (hi_32, lo_32)
  zigzag-delta-varint + per-band deflate. Honest measurement: 3.23× on 100KB
  native .omc, loses to plain zlib's 4.70× because the lo-band's pseudo-random
  deltas inflate under zigzag-varint. Code shipped, but rated negative; skip
  guards prevent it harming anything. New MCP tool omc_memory_compact_hbit.

- Axis 4-redo SBPE (OMCB): ★ first axis to actually beat zlib. Self-training
  greedy frequency BPE per body (512 merges by default), then ships
  zlib(merge_table) + zlib(token_stream). Measured 5.21× on 100KB native .omc
  vs 4.70× for zlib. The merge table travels inline (~1.5KB at 512 merges),
  amortizing for bodies ≥16KB. Round-trip verified end-to-end via MCP.
  New MCP tool omc_memory_compact_bpe.

Falsifications worth recording (all measured, none shipped as a default):
- PhiPC arithmetic codec with phi_pi_fib context routing: ceiling ~2.30×,
  ~half of zlib. Byte-Markov entropy bound is below LZ77 on text content.
- ONN-style Fibonacci-chunked parallel BPE + vocab union: 4.39× best,
  loses to single-pass global BPE (5.21×). Compression rewards global
  pattern frequency over local specialization — opposite of context
  compression where ONN wins.
- Substrate-weighted merge selection in BPE: decorative, no measurable lift.
  Pure greedy-frequency wins.

Magic-byte registry: OMCZ (zlib) | OMCT (substrate-tok) | OMCH (HBit)
                     | OMCD (delta) | OMCB (BPE — winner).

products/omc-memory-plus/README.md updated with the 5.21× headline and
the axis benchmark table.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: RandomCoder-lab