feat(erp): select shape-matched profiles and model window costs

[zzylol]

Why

Exact descriptor equality cannot reuse empirical sketch measurements across nearby production shapes, while window planning needs to compose atomic costs for recurrence, retention, panes, and sharing.

What

• Select the nearest compatible ERP by explicit distribution family, family-specific parameters, log-cardinality distance, and benchmark sufficiency.
• Compose update, merge, query, and retained-state counts for tumbling and pane-based windows.
• Keep the complete ERP architecture in one canonical design document.
• Preserve legal context-assigned fallback states in executable DAG validation.

How

erp_shape is extensible: {family, parameters, cardinality, benchmark_events}. Different families and unequal parameter schemas never interpolate. Event count is a minimum evidence gate rather than a distance axis. Eligible records are filtered by measured accuracy and ranked by composed workload cost.

Before this PR

ERP selection required exact distribution JSON and callers had to invent window operation multipliers. The DAG validator also rejected a legal shared raw fallback emitted by backend compilation.

After this PR

Deployments can match catalogued uniform, Zipf/discrete-power-law, continuous-power-law, normal, empirical, or future distribution families within explicit bounds. A miss remains observable so Hybrid callers can use Planner theoretical sizing or exact execution.

Verification

• cargo test -p asap-aware-mapping erp --lib — 6 passed.
• cargo test -p asap-types executable_dag — 2 passed before the shape-schema extension.
• ASAPQuery-backend ERP unit tests against 378ed16 — 8 passed.

Limitations and follow-up

Planner consumes family descriptors but does not itself fit arbitrary distributions. Runtime observation/fitting belongs to the backend; benchmark catalog generation belongs to sketch-bench. Burst behavior remains scenario provenance rather than an interpolation axis.

[zzylol]

Follow-up fix included: executable DAG validation now permits context-assigned maintenance/read states on shared raw fallback leaves. This was required for the backend full control-plane suite; Planner executable-DAG tests pass and the backend suite passes 716 library + 29 binary + 8 integration tests against commit 85cd1f7.