fix-mps-encoding-non-factorizing: requirements change

[jascal]

Summary

• Documents that the rung-1 hierarchical-polysemy MPS staircase shipped in add-mps-concept-encoding: §1, §3-§7 — hierarchical polysemy example, demo, docs #46 has Schmidt rank 2 but its Gram still factorizes as ⟨c_i|c_j⟩ = ∏_k cos((θ_{i,k} − θ_{j,k})/2) — identical to the rung-0 product-state Gram with the same angles. The four-tier hierarchy comes from angle design, not entanglement.
• Proposes a cross-coupled-by-sum encoding Ry(q0,a); CNOT; Ry(q1,a+b); CNOT; Ry(q2,b+c) that preserves Schmidt rank 2 and produces a non-factorized Gram (diff ≈ 0.32 vs. same-angle product-state).
• Sharpens two capability requirements: language (Hierarchical Polysemantic Example Pattern adds a non-factorization criterion) and compiler (MPS Concept Gram Matrix Analysis Helper accepts linear-combination angle expressions, option 2a).

This is a requirements-only change. No example, helper, demo, or docs are touched in this PR — those are the 29 tasks in tasks.md to be implemented in a follow-up PR.

Test plan

• openspec validate fix-mps-encoding-non-factorizing --strict passes
• openspec list shows the change as pending (0/29 tasks)
• Reviewer sanity-check: design.md §"Why this particular staircase factorizes" math intuition is correct
• Reviewer sanity-check: the three alternatives (A: cross-coupled-by-sum, B: brick-wall, C: γ-cross-coupled) and the recommendation rationale (A: symmetric cross-coupling, lowest framing churn, smallest helper rewrite) are sound
• Reviewer sanity-check: the post-mortem text in design.md §"Post-mortem entry in archived design" is the appropriate single edit to archived add-mps-concept-encoding/design.md

🤖 Generated with Claude Code