process: systemic premature-closure pattern — 7+ feature trackers closed before integration (orchestration audit evidence) #6836

Discovered during

User correctly pushed back during issue #4048 closure: the code I called "orphans" is not abandoned — it's completed, unit-tested, design-documented functionality whose tracker issues were closed before integration.

This is a process bug, not just individual tech debt.

Evidence

Each docstring explicitly cites a tracker issue. Each tracker is marked closed. None of the corresponding code is reachable from a production request path.

Plus #3228 (graph executor unification) — also CLOSED while code comments still reference it as "tracked in #3228" for unfinished work — see #6826.

Pattern

Issues are being closed when code is written and unit-tested, not when acceptance criteria for integration are met. The CLAUDE.md "Issue Closure Verification Gate (MANDATORY)" rule exists exactly to prevent this:

Verify each acceptance criterion: Is the feature/fix actually implemented? Does it work correctly?

"Working correctly" requires reaching production code, not just test code.

Why it matters

• ~3,900 LOC of unused completed features sit in the tree, contributing to god-class metrics, increasing build time, confusing readers about what's canonical
• Stakeholders who closed those issues think the features shipped — they didn't
• The original goals of the issues (e.g., "shared memory between parallel workflow agents" enhancement: no shared memory between parallel workflow agents #3019) are not actually delivered to users

Recommendation

Process changes

1. Add an "integrated" acceptance criterion as a default in CLAUDE.md "Issue Closure Verification Gate":

   Closure requires: (a) code merged, (b) at least one production code path imports the new module, (c) integration test exists.

2. Closure audit: when closing a feature issue, run grep -rn 'from <new_module>' autobot-backend --include="*.py" | grep -v _test — must return ≥1 hit.

Tactical follow-up

1. Reopen trackers Enhancement: Q-Learning reinforcement router for agent task routing #2092, feat: Agent mesh topology — dynamic DAG with Hebbian evolution (Neural Mesh RAG Phase 5) #2137, feat: Topology-aware routing + agent specialization (Neural Mesh RAG Phase 5) #2138, Enhancement: Workflow resume-from-failure and step-level error handlers #2154, enhancement: no shared memory between parallel workflow agents #3019, feat(backend): custom success criteria definitions for workflow orchestrator #3293, plus Code Smell: Refactor 178 god class patterns #381 sub-trackers — none meet integration acceptance criteria.
2. The 14 follow-up issues from refactor(backend): complete orchestrator consolidation (7→4 target) #4048 closure (discovery(orchestration): wire in or retire abandoned causal subsystem (5 files, 1,722 LOC, 0 production callers) #6816–discovery(orchestration): subagent_dispatcher.py is test-only, lives in services/ instead of orchestration/ (layering violation) #6822, discovery(orchestration): #3228 closed prematurely — executor fragmentation persists (5 files, 3,223 LOC, no unified strategy) #6826–discovery(orchestration): low-utilization composability — SuccessCriteriaEvaluator + CollaborationCoordinator are single-caller modules #6832) collectively represent the wire-in work. They are the path to making each reopened tracker actually deliverable.

Acceptance criteria

• CLAUDE.md updated with explicit integration check in closure gate
• Reopen the 7 prematurely-closed feature trackers OR document why each was correctly closed despite zero production callers
• Each of discovery(orchestration): wire in or retire abandoned causal subsystem (5 files, 1,722 LOC, 0 production callers) #6816, discovery(orchestration): wire in 4 orphaned modules — workflow_planner, workflow_documentation, workflow_memory, agent_registry (~1,260 LOC, 0 production callers) #6820, discovery(agents): wire in or retire topology modules — AgentTopology + TopologyAwareRouter (336 LOC, 0 callers, #2137/#2138) #6821, discovery(agents): RLRouter (Issue #2092) infrastructure-complete but unreachable in production paths (354 LOC) #6829, discovery(orchestration): low-utilization composability — SuccessCriteriaEvaluator + CollaborationCoordinator are single-caller modules #6832 cross-links its corresponding reopened tracker

Priority

P1 — process correction. Blocks credibility of "closed" issue counts as a measure of delivered features.