Remap all v1 templates to the consolidated v3.0.0 skill names (hold for release day)

[cafzal]

What

Prepares the templates repo for the RAI skills v3.0.0 release, which consolidates 15 skills into 11. Every skill invocation and prose mention under v1/ — runbook slash commands, README notes, script comments — moves to the new names.

Do not merge until the v3.0.0 skills release ships to the public repo. The old skill names keep working until then; the new ones don't exist for customers yet. This PR ships the same day as the release.

Name map

Old	New
rai-querying, rai-rules-authoring, rai-pyrel-coding	rai-pyrel
rai-build-starter-ontology, rai-ontology-design	rai-ontology
rai-prescriptive-problem-formulation	rai-prescriptive-problem
rai-prescriptive-solver-management (solver selection, pre-solve)	rai-prescriptive-problem
rai-prescriptive-solver-management (solve execution, scenarios, diagnostics)	rai-prescriptive-results
rai-prescriptive-results-interpretation	rai-prescriptive-results

Unchanged: rai-setup, rai-discovery, rai-graph-analysis, rai-predictive-modeling, rai-predictive-training, rai-deployment, rai-health.

Beyond the rename

• Combined invocations collapse: problem-formulation + solver-management steps become a single /rai-prescriptive-problem step; solver-management + results-interpretation scenario steps become a single /rai-prescriptive-results step.
• In commercial_underwriting and shipment_compliance the merge left a rules step immediately followed by a query step that only read the derived flags back; each pair collapses into one /rai-pyrel step and the surrounding prose is reworded.
• ASCII stage diagrams re-aligned where shorter names shifted columns (defect_root_cause, entity_resolution, telco_network_recovery).
• Skill-content references verified against the consolidated tree: rai-prescriptive-problem/scenario-analysis.md (cited in supply_chain_resilience.py), rai-pyrel data-loading guidance (cited in two READMEs), TIME_LIMIT-is-signal guidance now under rai-prescriptive-results.

Validation

Applied the dev-templates-review checklist to the diff: chain ASCII stage labels carry the new skill names with columns re-aligned, merged prompts stay question-shaped with definitions inline, Responses stay single-paragraph, no step-number cross-references introduced, headline numbers untouched. Touched scripts pass py_compile and ruff.

Paste-test sweep results

Ran the full paste-test sweep once the consolidated v3.0.0 skills library locked (claude/skill-streamlining @ 88f7121). Protocol: fresh Sonnet agents, Snowflake engines, one runbook prompt delivered per turn in sequence — no read-ahead, no injected model/solver names, schemas, or expected answers; the only context beyond each prompt's own verbatim text is the runbook's own run-in-order preface. Scoped as a regression test: each template ran through its last step that touches a modified skill (rai-ontology, rai-pyrel, rai-prescriptive-problem, rai-prescriptive-results), including genuine prerequisite steps, skipping trailing steps that only exercise unchanged skills (pure graph/predictive/discovery tails with nothing downstream).

46 of 48 templates reached a full PASS, with zero hard failures anywhere in the sweep. The regression unit is the runbook step, not the template: every one of the 187 modified-skill steps that ran passed — 0 failures, 0 regressions (187/187), each objective/count cross-checked against the solver or an independent numpy/pandas oracle. Per skill: rai-ontology 49/49, rai-pyrel 63/63, rai-prescriptive-problem 39/39, rai-prescriptive-results 36/36. (A further 10 modified-skill steps in the corpus went unrun — not failed — almost all the terminal/build steps of the two GNN-blocked templates.) The remaining 2 templates (subscriber_retention, telco_network_recovery) are environment-scope-limited: their final modified-skill step genuinely requires a GNN prediction as input, this environment's relationalai build lacks the [gnn] runtime extra, and neither ships a precomputed-prediction fallback CSV — an environment gap, not a template or skill defect. (Several templates' background agents were killed twice by session interrupts mid-run and relaunched from scratch; the final PASS numbers are from completed clean runs.) Zero PASS templates required a skill-name fix — the remap itself holds up under real execution.

Runbook-prompt issues found were fixed directly in this PR (8 files): named the restricted/marginal industry codes in commercial_underwriting; made the persist step explicit in bom-reachability; quantified the flow-conservation tolerance in money_laundering_motif_detection; de-overloaded "relationships" in fraud-detection; named exact cost weights in defect_root_cause (previously unspecified weights could name different real-world root-cause factors on different runs); named the SPOF disambiguation criterion and allowed for ties in two memory_supply_allocation prompts; and stated the discount-monotonicity rule explicitly in retail_planning (matching retail_markdown's existing wording). Two other suspected issues were investigated and retracted after checking the real template scripts/data: an apparent memory_supply_allocation baseline infeasibility turned out to be a test-agent implementation slip (mixed up which Dependency-edge role feeds which property, not a runbook or skill defect), and a suspected retail_planning Article/PlanningArticle id mismatch turned out to be a potential Snowflake data-provisioning question, not a prompt-wording gap — flagging that one to the data team rather than editing the prompt.

Skill-content findings (not actionable in this repo — reported to the skills team): two correctness-class findings worth prioritizing before release — in-model aggregation over Scenario-indexed decision variables can silently return a wrong total with no error (demand_planning_temporal), and selecting two aggregate-derived properties together can silently return a different value for one of them than selecting it alone (fraud-detection). Also: a NaN in any column of a multi-column model.data(df) silently drops the whole row from entity creation even when only a non-null column is referenced (datacenter_compute_allocation, supply_chain_resilience), and a Python-level variable/label matching a concept's attribute name can silently mint a real server-side property when implicit_properties is on (datacenter_compute_allocation). Plus a handful of lower-severity friction points (.to_schema() column access, solve_for(where=...) needing an inline literal, pre-solve decision-variable queries returning 0 rows, .alias() Property-vs-Relationship behavior, model.union() branch-arity mismatch) where agents self-corrected from the skill's own reference material. Full writeup: dev_temp/v3_skills_sweep_skill_issues.md.

Skill-file coverage (from real Read calls across all agent transcripts, not estimated): rai-ontology 9/18, rai-pyrel 16/33, rai-prescriptive-problem 35/55, rai-prescriptive-results 7/20 (thinnest — untouched files cluster around diagnostics/conflict-IIS/failure-taxonomy and Pareto/multi-objective paths that no template in this set happened to exercise). The per-file breakdown of which reference/example files were exercised vs. never opened is captured on the skills PR (RelationalAI/rai-agent-evals#157) as input for a follow-on skill-streamlining pass.