Rebuild the Goals 1–5 pipeline at scale (Group A)

[duckyquang]

Context

The v0 prototype (geml_experiments) completed Goals 1–5R on a 10k corpus (262/262 tests green). It proved the measurement methodology but is not a template for the production pipeline.

⚠️ Do not hand the prototype repo wholesale to a coding agent — it will copy the structure, which does not scale. Use it to understand what was measured and why; reimplement from scratch here, with thorough human + agent code review.

Scope (rebuild at 10–100× corpus scale)

• Expression generator with 3R discipline (depth mixture, dedup, triviality caps, byte-reproducibility)
• Bi-directional SymPy/LaTeX ↔ AST ↔ pure-EML converters (official EML compiler, no abbreviations, no hidden derived leaves)
• Expansion study: α distributions vs the 1 + log_4L(K) threshold, re-measured on the new corpus
• Exact structural DAG compression (lossless control channel)
• E-graph compression: safe + positive-real modes, EML-aware extractor (kept as optional canonicalization pre-pass)
• Frequent-motif compression with verified lossless reconstruction (primary practical channel; learned motif selection stays OFF the critical path)
• Metrics, tests, CI green on fresh clone; per-goal docs written as work lands (no end-of-project docs crunch)

Compute

H100 hours if available; otherwise RTX 5090 over SSH (sufficient for generation + motif/macro extraction).

See README §4–5 for the v0 numbers each stage must reproduce or beat.