[zzylol]

Why

ASAP primitives expand the search space and create cross-query cost dependencies. This proposal explains compact candidate generation and joint cost-based selection.

What

Adds a concise memo/MILP design and a separate window/materialization proposal. The main document uses connected explanatory prose and design rationale, with traditional memo rule examples followed by P95/P99 through ASAPPlanner stages. It cites Graefe’s Cascades paper and distinguishes demand-driven exploration from separate exhaustive phases.

The proposed representation has groups and physical candidates. Multiple candidates from one group may be selected; consumers connect directly to selected candidates. Shared state is represented by memo groups, without separate physical-instance slots. Distinct execution scopes or incompatible input bindings require distinct generated candidates.

The MILP defines candidate selection, compatible connections, query coverage, dependency support, acyclicity, and cost/accuracy/resource requirements. Optimality is restricted to the generated inventory and configured models. Proposed behavior is distinguished from existing implementation.

Key code interfaces added / modified

None (documentation only).

Before this PR

No unified proposal connecting memo expansion, ASAP-aware shared state, window/materialization choices, and joint selection.

After this PR

P95 and P99 remain separate result groups and can connect to one shared KLL candidate. Several KLL configurations may also be selected for different consumers. The design explains this with examples and a compact MILP formulation.

Verification

• git diff --check passed.
• Checked section numbering, Variables/Constraints subsections, local links, code fences, and candidate terminology.
• Main document is approximately 2,500 words, with examples preceding the formal model.
• Documentation only; no runtime tests run.