Section 04: Reviewer Wiring + Mechanical Enforcement

Status: Complete Goal: Close the “training-only enforcement” gap Opencode flagged during /tp-help. Reviewers recognize the new PLAN_ROUTING_DRIFT category by virtue of the grounding preamble reading impl-hygiene.md (§02’s work) — but training-only failed to prevent 22 re-route plans accumulating. Mechanical gate at plan_corpus check --strict-recon time catches drift before reviewers do.

Success Criteria:

tp_agent_prompt.md carries a PLAN_ROUTING_DRIFT sample finding in its return-schema examples (connects to mission criterion: “Reviewers flag new peer plans without Fallback-R justification as CRITICAL”)
plan_corpus check --strict-recon fires an ERROR (exit 1) for plans/<name>/index.md with reroute: true AND missing routing_justification:
Synthetic sibling /tmp/test-sibling/ with reroute: true but no justification triggers PLAN_ROUTING_DRIFT
/review-plan role: hybrid — no --called-by-orchestrator branch needed (the finding category is constant)

Context: All three reviewers during /tp-help agreed: finding categories are schema-definition-time, not invocation-time. Severity is set in impl-hygiene.md (landed in §02); reviewers pick it up from grounding. The gap is mechanical enforcement — the 2026-04-24 sibling-fix-BUG ban relied on reviewer training only, and 22 re-route plans accumulated anyway. This section adds the belt-and-suspenders plan_corpus validator gate.

Reference implementations:

tp_agent_prompt.md existing rule_violated: examples — format to mirror for the new category.
scripts/plan_corpus/schemas.py TprInfo validator — precedent for “field X is required when field Y is true” conditional validation.
Sibling fix-BUG ban (2026-04-24) — precedent for how training-only enforcement fails at scale; motivation for the mechanical gate.

Depends on: Section 02 (PLAN_ROUTING_DRIFT category must exist in impl-hygiene.md before reviewer example can reference it), Section 03 (plan_routing.py must exist to verify the synthetic sibling case in 04.3).

Intelligence Reconnaissance

Queries run 2026-04-24:

grep -c "rule_violated:" .claude/skills/tpr-review/tp_agent_prompt.md — counts existing category examples.
grep -n "TprInfo\|@dataclass" scripts/plan_corpus/schemas.py — locates validator pattern precedent.
scripts/intel-query.sh --human callers "classify_bug_exclusion" --repo ori — confirms bug_markers.py is imported by two scanners, validating the SSOT pattern we’re mirroring.

Results summary (≤500 chars) [ori]: tp_agent_prompt.md return-schema already has rule_violated: example lines for existing categories (INVERTED-TDD, LEAK, DRIFT); adding PLAN_ROUTING_DRIFT is a one-line addition. scripts/plan_corpus/schemas.py uses @dataclass(frozen=True) validators (e.g., TprInfo for third_party_review: nested block validation) — same pattern applies for routing_justification: conditional-required validation.

See .claude/skills/query-intel/compose-intel-summary.md for the full query protocol.

04.1 Add PLAN_ROUTING_DRIFT example to tp_agent_prompt.md

File(s): .claude/skills/tpr-review/tp_agent_prompt.md

Locate the return-schema / rule_violated: example block and add a new example finding.

grep -n "rule_violated:" .claude/skills/tpr-review/tp_agent_prompt.md → identify insertion point

Add a sample finding example near existing category examples:

- id: "EXAMPLE-1"
  severity: "critical"
  path: "plans/new-cross-cutting-work/index.md"
  line: 2
  title: "Sibling plan created without routing_justification — PLAN_ROUTING_DRIFT"
  evidence: "plans/new-cross-cutting-work/index.md has reroute: true but no routing_justification: field; scope touches ori_arc + ori_llvm with primary-by-count mapping to section-21A-llvm (should have been a subsection, not a peer directory)"
  rule_violated: ".claude/rules/impl-hygiene.md §Finding Categories — PLAN_ROUTING_DRIFT (Critical)"
  recommended_fix: "Run `python3 scripts/plan_routing.py --scope ori_arc,ori_llvm` to get the mechanical routing decision; either (a) migrate the plan's content into a subsection of section-21A-llvm per the helper's output, or (b) add routing_justification: explaining why the plan cannot live as a subsection (e.g., if primary-by-count returns a tie)."

Run prose-lint: python3 scripts/prose-lint.py .claude/skills/tpr-review/tp_agent_prompt.md exits 0
Verify grep: grep -c "PLAN_ROUTING_DRIFT" .claude/skills/tpr-review/tp_agent_prompt.md returns at least 1
Subsection close-out (04.1) — MANDATORY before starting 04.2:
- Example added and grep-verified; prose-lint green
- Update this subsection’s status to complete
- Repo hygiene check — compiler_repo/diagnostics/repo-hygiene.sh --check

04.2 Add mechanical plan_corpus check —strict-recon rule

File(s): scripts/plan_corpus/schemas.py (validator addition), scripts/plan_corpus/bug_validators.py or new scripts/plan_corpus/plan_validators.py

Conditional-required validation: routing_justification: str must be present when reroute: true; must be absent when reroute: false or reroute is None.

Decide placement: extend existing bug_validators.py (if generic enough) OR create scripts/plan_corpus/plan_validators.py for plan-level checks; prefer the latter for SRP

Implement validation function:

def validate_reroute_routing_consistency(plan_index_path: Path, fm: dict) -> list[str]:
    """Returns list of diagnostic strings; empty = clean."""
    diagnostics = []
    reroute = fm.get("reroute")
    justification = fm.get("routing_justification")

    if reroute is True and not justification:
        diagnostics.append(
            f"{plan_index_path}: reroute: true requires routing_justification: (see CLAUDE.md §Plan Routing — Mechanical)"
        )
    if reroute is not True and justification:
        diagnostics.append(
            f"{plan_index_path}: routing_justification: is only permitted when reroute: true (currently reroute={reroute})"
        )
    return diagnostics

Wire the validator into plan_corpus check --strict-recon dispatch so findings surface as ERRORs (exit 1), not WARNINGs
Verify by running python -m scripts.plan_corpus check --strict-recon plans/ on the current tree — should emit findings for any existing re-route plan lacking routing_justification:. If findings fire on current state, that’s expected — §05 migration clears them. For now, document which plans trip the gate in this section’s Manifest Appendix
Add unit test: fixture plan with reroute: true + missing routing_justification: → validator returns 1 diagnostic; fixture with both present → 0 diagnostics; fixture with routing_justification: but reroute: false → 1 diagnostic (forbidden)
Subsection close-out (04.2) — MANDATORY before starting 04.3:
- Validator added + unit-tested
- plan_corpus check --strict-recon surfaces PLAN_ROUTING_DRIFT ERRORs on bad fixtures and current non-compliant plans
- Update this subsection’s status to complete
- Repo hygiene check — compiler_repo/diagnostics/repo-hygiene.sh --check

04.3 Dry-run verification on synthetic sibling plan

File(s): /tmp/test-sibling-plan/ (scratch only — not committed)

End-to-end verification: the full reviewer chain catches the drift mechanically.

Create synthetic fixture:

mkdir -p /tmp/test-sibling-plan
cat > /tmp/test-sibling-plan/index.md <<'EOF'
---
name: "Test Sibling"
full_name: "Test Sibling Plan (Should Fire Drift)"
status: queued
reroute: true
# routing_justification: INTENTIONALLY OMITTED to trigger drift
---
# Test Sibling Plan
EOF
cat > /tmp/test-sibling-plan/00-overview.md <<'EOF'
---
plan: "test-sibling"
title: "Test Sibling"
status: not-started
---
# Test Sibling
## Mission
This is a test fixture that should trigger PLAN_ROUTING_DRIFT.
EOF

Run python -m scripts.plan_corpus check --strict-recon /tmp/test-sibling-plan/ → expect exit 1 with PLAN_ROUTING_DRIFT message naming routing_justification requirement
Run /tpr-review --help-mode --max-rounds=1 --prompt-file=<objective citing /tmp/test-sibling-plan/ as the review target> — expect at least one reviewer to flag PLAN_ROUTING_DRIFT Critical (this is the dry-run verification per /create-plan Phase 5 cohesion pattern)
Cleanup: rm -rf /tmp/test-sibling-plan/
Record dry-run outcome in this section’s Manifest Appendix Part E (add below at end of section): which reviewer(s) fired the finding, at which severity
Subsection close-out (04.3) — MANDATORY before starting 04.N:
- Dry-run confirms both mechanical gate AND reviewer-training gate fire on the synthetic case
- Update this subsection’s status to complete
- Repo hygiene check — compiler_repo/diagnostics/repo-hygiene.sh --check

04.N Completion Checklist

Exit Criteria: Reviewer example published; mechanical gate rejects malformed siblings at plan_corpus check --strict-recon; synthetic fixture dry-run fires PLAN_ROUTING_DRIFT Critical from both enforcement paths.