docs: open-problems thesis + re-derived B→E roadmap (space-bounded genomics as a memory contract)

[logannye]

Summary

Captures the first-principles re-derivation of Rosalind's research thesis and the B→E roadmap, optimizing for the builders/researchers who need its unique capabilities.

New: docs/OPEN_PROBLEMS.md — research-framed strategic doc:

• Thesis: memory as a declared, predictable, never-refusing, verifiable contract across the whole lifecycle, with the square-root-space machinery (Williams 2025 / Cook–Mertz 2024; bound O(√(t·log log t))) as the continuous space/time knob a declared RAM budget selects.
• Open problem: not "can we hit √-space" but "for which genomics computations can we realize the full space/time curve with tolerable time overhead?"
• Decisive lens: the framework shrinks computation intermediate state, not inputs — so the beachhead is sublinear-space FM-index/SA construction (intermediate-state-bound, one-time/offline, on the critical path, erases the O(reference) build-RAM caveat). Input-bound problems (cohort calling, pangenome graphs) are framed honestly as needing orthogonal techniques.
• Contract design: rosalind plan envelope + curve; honor-or-refuse → graceful degradation; lifecycle-wide budget; verifiable receipts.

Realigns …/rosalind-target-architecture.md:

• §1 — sharpened core bet (+ pointer to OPEN_PROBLEMS).
• §8 — theory layer KEPT and repurposed as load-bearing (the √t-family is the budget knob), superseding the earlier "feature-gate the demo" decision.
• §9 — re-derived sequencing: B genome-scale kernel → C memory-as-verifiable-contract → D sublinear-space construction (beachhead) → E reach + programmable substrate. Performance/breadth deliberately follow the unique capability.

Test plan

• Docs only; no code changes. B→E roadmap consistent across both docs; bound and theory-layer decision consistent.

🤖 Generated with Claude Code