vol2: ChangeSet delta not passed to EvaluationEngine — engine always does full rebuild

[mdproctor]

Problem

RuleBaseModifier.apply(changeSet) builds the delta (added/removed rules) but throws it away after accumulating into UnitDescriptor. The engine receives the full UnitDescriptor — all rules ever added — not the delta. Every changeset triggers a full Router + BetaProcessor rebuild.

Current flow

RuleBaseModifier.apply(changeSet)
  → unitDescriptor.addRule(rule)   // all rules accumulated
  → compiledEngine = null           // invalidate → full rebuild next createUnit()
createUnit()
  → engine.compile(unitDescriptor, rete)  // sees ALL rules, full rebuild

Correct design

The ChangeSet IS the diff. RuleBaseModifier.apply() is the right seam to coordinate incremental engine updates — NOT a new method on RuleBase (which stays minimal, see indirection principle).

RuleBaseModifier.apply(changeSet)
  → if compiledEngine exists:
      engine.patch(compiledEngine, added, removed, rete)
      // Router gets new processors for added rules, removes for removed rules
  → else:
      compiledEngine = engine.compile(unitDescriptor, rete)  // initial build only

Design context

RuleBase intentionally exposes no mutation methods — all changes go through RuleBaseModifier. This is the vol2 "one degree of separation" principle: RuleBaseModifier owns changeset coordination and engine patching; RuleBase holds compiled state only. This avoids the vol1 mistake of coupling session creation directly to KieBase, which made the API impossible to refactor.

Also connects to

• compile() naming concern — compile() implies a full build; an incremental operation needs a different verb (patch()? update()?)
• Proper incremental NotNode/ExistsNode evaluation (vol2: proper NotNode/ExistsNode Rete evaluation — replace lambda scope checking in head processors #6713) — incremental patching requires knowing which processor chains to update

Refs #6712, #6718