Add analyst runbooks for 36 v1 templates

[cafzal]

What

Adds paste-testable runbook.md walkthroughs to every v1 template that lacked one (36 templates), bringing the whole v1 catalog to parity with the existing runbooks (energy_grid_planning, telco_network_recovery, it-dependency-mapping, …).

Each runbook is the analyst-facing bridge between the ontology, the script, and a Director-of-Analytics-style user: a brief problem statement, a distilled chain diagram, and an ordered set of question-shaped prompts (each naming its /rai-* skill) with the response the chain produces. Single-reasoner templates use the compact format; multi-reasoner templates use the staged fenced format with a /rai-discovery routing step.

Coverage (36)

• Prescriptive (21): ad_spend_allocation, campaign_roi, cell_tower_coverage, cicd_runner_allocation, demand_planning_temporal, diet, factory_production, financial_index_replication, hospital_staffing, money_laundering_motif_detection, network_flow_planning, pod_placement, product_configurator, production_planning, retail_markdown, shift_assignment, sprint_scheduling, supplier_reliability, supply_chain_transport, traveling_salesman, underwriting_audit, water_allocation
• Graph (3): disease-outbreak-prevention, humanitarian-aid-supply-chain, wildlife-conservation-network
• Rules (2): commercial_underwriting, shipment_compliance
• Predictive / GNN (2): demand_forecasting, smoker_status_prediction
• Multi-reasoner (7): book_slate_recommendation, fraud-detection, patient_cohort_recruitment, planogram_optimization, retail_planning, subscriber_retention, warehouse_allocation

Verification (folder-only, on Snowflake)

31 of the 36 runbooks were independently reproduced by a fresh, skill-loaded agent that had not seen the script. The test is deliberately clean-room: the agent gets only the template's data/, the runbook's pasteable prompts, and the named /rai-* skills — with no injected CSV headers, solver, problem type, seeds, or expected answers, so "matches the runbook" can't be circular. Every run executes on Snowflake (the engine customers use), and the headline numbers in each runbook are that run's real output.

The remaining 5 GNN-training templates (demand_forecasting, smoker_status_prediction, fraud-detection, subscriber_retention, retail_planning) are accepted via ground-truth script runs rather than the clean-room paste-test, since they require GPU training.

Folder-only validation corrected 11 runbooks whose stated figures depended on context not present in the pasteable prompts — a normalization convention (disease-outbreak, humanitarian: raw vs n−1-normalized weighted degree), an underspecified rule (underwriting_audit: pin the review rule so the deliberate-bug counts reproduce), or a solver-path-/training-dependent objective (ad_spend_allocation, demand_planning_temporal, financial_index_replication, hospital_staffing, pod_placement, planogram_optimization, supply_chain_transport, book_slate_recommendation). Each was softened or pinned to what an independent run reproduces.

Also brings six pre-existing runbooks up to the runbook standard (supply_chain_resilience, memory_supply_allocation, portfolio_balancing, datacenter_compute_allocation, bom-reachability, cybersecurity-attack-paths).

README fixes (also in this PR)

A /dev-templates-review pass across all v1 READMEs surfaced seven unambiguous, customer-facing defects, fixed here:

• Broken Quickstart download URLs — private.relational.ai → docs.relational.ai (cell_tower_coverage, financial_index_replication).
• Quickstart unzip/cd names that didn't match the hyphenated .zip, breaking copy-paste (disease-outbreak-prevention, humanitarian-aid-supply-chain, wildlife-conservation-network).
• Setup-breaking GRANT typo — FAVORITA.EXPERIMENTS → FAVORITA_MINI.EXPERIMENTS (demand_forecasting).
• Stray </content></invoke> tool markup rendering as literal text (memory_supply_allocation).

Deferred to a follow-up (not in this PR)

The same review flagged two larger bodies of work, kept out to keep this PR focused on runbooks:

• Portfolio-wide README standardization (~45 READMEs): remove the body H1 (the gallery renders the title from front matter), add ## Model overview / ## Learn more / ## Support, convert ## Customize this template to the H3 subheadings, split ## Prerequisites into Access/Tools, trim oversized Quickstart "Expected output" blocks. The bulk of the portfolio predates this standard.
• README "Expected output" vs script/runbook reconciliation (~12 templates): several READMEs print stale numbers that disagree with the shipped script (e.g. supply_chain_transport $5,080 → $1,444, retail_markdown $18,432.50 → $23,374.65, production_planning's stale 3-solve loop output, diet's basket, demand_planning_temporal's week count). These need per-script reruns.
• Two script-level bugs seen while running templates end-to-end: fraud-detection ships FRAUD_DB where the data + EXPERIMENTS schema is FRAUD_DETECTION; hospital_staffing's script prints $1,728 where the README gives $1,116.