Revise the proposal under the agreed claim discipline

[duckyquang]

The proposal document still carries claims the v0 data contradicts. Rewrite it to match the README's reframing:

• "EML shrinks math" → hypothesis: raw EML is ~11–12× larger (median α ≈ 11.4 vs threshold ≈ 1.56); compression makes it trainable; generalization benefit is what's under test
• "GPT-4-level competency at a fraction of parameters" → "competitive accuracy with fewer parameters on a bounded algebraic domain" (tested explicitly in Goal 11.2)
• "Universal / zero-shot theorem proving" → "verified equational rewrite-proofs on a bounded domain"
• Remove "equivalence as graph isomorphism" — equivalence is semantic, not structural
• Remove/demote the "machine code for AI" framing to a speculative future-work remark
• Promote baselines in the methodology: AST-DAG + compute-matched transformer are essential controls — without matched-compression AST, nothing is attributable to EML