Dataflow: Atomic rule-generation rebuild and swap #185

Problem

Rule/schema replacement must not expose output from mixed generations while persistent graphs rebuild.

Evidence

Source: 5c1f313379a5cb2995a43a66ddba0b5cc9798b7f (inspected 2026-10-03). These are source findings; the acceptance scenarios below still need implementation and execution.

• src/protocol/subscription/reevaluate.rs:76: The current standing-query adapter recomputes results.
• src/storage_engine/mod.rs:2350: Existing incremental-engine activation/shadow infrastructure is the starting point.

Acceptance criteria

• Build a new versioned graph off the serving path and swap at a committed frontier; failed compilation preserves the complete old generation.
• Define cancellation/drop invalidation and reset semantics for subscribers; keep compile/cache ownership outside the handler monolith.
• Test concurrent writes, schema arity changes and late old-generation completions; measure rebuild pause, RSS and warm-query latency.
• Compare every visible supported revision through Testing: revision-aware differential correctness oracle #167; pass Performance: baseline and mandatory pre-PR regression pipeline #163 for query/write/delta latency, throughput and memory before default adoption.

Delivery

Plan item: D06. R4 - Persistent incremental dataflow.

Dependencies / related work: #184, #170.

Contributes to #143; completing this slice alone does not close the broader issue.