docs: add cross-CLI learning control-plane implementation plan + DAG

docs/demos/council-next-steps-demo.md

@@ -0,0 +1,106 @@
+# Demo Runbook: /council decides LACP next steps
+
+Objective
+Demonstrate that LACP can use council-style deliberation + measurable eval evidence to decide safe next steps for cross-CLI self-learning.
+
+Audience takeaway
+- LACP is not guessing; it uses structured deliberation with dissent.
+- Learning is gated by metrics and rollback safety.
+- Next actions are concrete and reversible.
+
+Duration
+20-30 minutes
+
+---
+
+## Prep (5 minutes)
+
+Required
+- repo checked out at `main`
+- `gh` authenticated
+- baseline and treatment eval fixtures available
+
+Set safety defaults
+```bash
+export LACP_LEARNING_ENABLED=1
+export LACP_LEARNING_MODE=shadow
+```
+
+Show current status
+```bash
+git rev-parse --short HEAD
+bin/lacp-status-report --json | jq '.intervention_rate_kpi? // .'
+```
+
+---
+
+## Part 1: Council deliberation (8-10 minutes)
+
+Prompt frame
+- Decision: What should LACP build in the next 7 days to become a safe self-learning layer across CLIs?
+- Constraints: no routing mutation in shadow, proven rollback, measurable uplift.
+- Success criteria: >=10% completion uplift, zero sev-1, rollback <5m.
+
+Run council process
+1) Strategist seat (control-plane and sequencing)
+2) Pragmatic shipper seat (smallest shippable slices + tests)
+3) Stoic risk seat (threats, containment, rollback)
+4) Synthesizer with anti-convergence and dissent preservation
+
+Expected output artifact
+- A/B/C option matrix
+- dissent summary
+- recommended path + fallback trigger
+- immediate 3-step plan
+
+---
+
+## Part 2: Before/after evidence (8-10 minutes)
+
+Run baseline vs treatment replay
+```bash
+bin/lacp-learn-eval --dataset data/benchmarks/learning-loop-fixtures.json --json | tee /tmp/lacp-learn-eval.json
+jq '{success_delta, retry_delta, intervention_delta, decision}' /tmp/lacp-learn-eval.json
+```
+
+Show guardrails
+```bash
+bin/lacp-provenance verify --json | jq
+bin/lacp-learn-rollback --last --dry-run --json | jq
+```
+
+Must show live
+- route parity proof for off vs shadow
+- provenance verification pass
+- rollback dry-run details
+
+---
+
+## Part 3: Decision gate (5 minutes)
+
+Go/No-go criteria
+- GO if:
+  - completion uplift >=10% on deterministic replay
+  - no shadow-mode route mutation
+  - rollback proof <5m
+  - no unresolved critical risk
+- NO-GO if any critical gate fails
+
+Fallback trigger
+- If uplift <10% after two weekly cycles, narrow scope to outcome-dense Learning Packs.
+
+Final call format
+- Decision: GO to next ring or NO-GO remain in shadow
+- Why: 3 bullets (evidence-backed)
+- Dissent: what remains unresolved
+- Next 72h tasks: exact commands/owners
+
+---
+
+## Demo checklist
+
+- [ ] Council artifacts captured
+- [ ] Eval artifact attached
+- [ ] Provenance verification artifact attached
+- [ ] Rollback proof attached
+- [ ] Final call documented in PR/issue thread