perf: evaluate adaptive K-fusion parallel dispatch for low-K recursive strata

[justinjoy]

Background

Issue #741 records a negative finding from PR #731: lowering WL_KFUSION_MIN_PARALLEL_K from 4 to 2 regressed CRDT W=8 by +10%. The fixed lower threshold made low-K recursive strata pay parallel-dispatch overhead across many iterations.

Current behavior remains:

• Plan generation emits WL_PLAN_OP_K_FUSION for recursive strata with K >= 2.
• Runtime parallel dispatch is gated by WL_KFUSION_MIN_PARALLEL_K == 4 in wirelog/columnar/ops.c.
• K=2 and K=3 still execute through the serial K-fusion evaluator.

Problem

A static threshold is too blunt. CRDT has thousands of recursive iterations where K=2 parallel overhead compounds. Other workloads may have different iteration counts and relation sizes. The failed experiment also showed that adding a stratum_is_recursive bit to wl_plan_op_k_fusion_t would be redundant, because K-fusion is already generated inside recursive strata.

Proposed direction

Investigate an adaptive runtime heuristic instead of a fixed threshold change, for example:

• Learn per-stratum or per-K-fusion-op whether parallel dispatch pays off after the first N iterations.
• Include relation sizes, live branch count, worker count, and observed dispatch/fill time in the decision.
• Keep K=2/K=3 serial by default until the learned signal justifies parallel execution.
• Preserve the existing fixed threshold as a conservative fallback.

Acceptance criteria

• Do not lower WL_KFUSION_MIN_PARALLEL_K globally from 4 to 2 without an adaptive guard.
• Do not add a redundant stratum_is_recursive field to wl_plan_op_k_fusion_t; planner context already guarantees recursive origin for K-fusion.
• Validate CRDT W=8 does not regress beyond the existing performance fence.
• Validate DDISASM and at least one additional workload do not regress beyond +/-2%.
• Document the heuristic and its fallback behavior in docs/THREADING.md if implemented.

References

• Follow-up to do-not-retry: lower WL_KFUSION_MIN_PARALLEL_K from 4 (CRDT W=8 +10%) #741
• PR perf(CRDT): leading-key compactor cache + perf-suite gate #731, "What was attempted but did not ship"
• wirelog/columnar/ops.c::WL_KFUSION_MIN_PARALLEL_K
• wirelog/exec_plan_gen.c::rewrite_multiway_delta
• Risk C4: Nightly perf gate continue-on-error #707 for release perf-gate policy context
• Epic: v0.42 Semantics Closure #682 v0.42 semantics epic