Trim Step 1 (build ontology) prompts to need-to-know

[cafzal]

Why

The Step 1 prompts in each template's runbook.md enumerated every CSV in data/ (e.g. "…from the eight CSVs in data/: cell_towers, network_equipment, …"). That's noise — /rai-build-starter-ontology introspects data/ and picks concept shapes itself. The prompts read more like an instruction manual than what a user would type.

This PR trims them to need-to-know only: the destination (data/) and any modeling hint that's non-obvious from CSV inspection alone.

Changes

Template	Hint kept	Rationale
telco_network_recovery	composite-key time-series (date+region)	not visible from headers; agent could synthesize a surrogate id
energy_grid_planning	(bare trim)	every CSV maps 1:1 to a Concept, no surprises
supply_chain_resilience	(bare trim)	same
machine_maintenance	introduce a Period Concept (1..N from period column)	Period is synthesized from an integer column, not its own CSV
portfolio_balancing	binary-property covariance (long-form (stock_i, stock_j, value)) + promote Sector to a Concept	covariance is naturally a binary property on Stock; sector enums need to be Concepts so downstream rules can aggregate

Test plan

• Static simulation — per-template subagents loaded the rai-build-starter-ontology skill, inspected CSV headers, and wrote out the PyRel ontology code an agent would produce from the trimmed prompt. All five concept sets match canonical.
• Live canonical run — each template's main script (which builds the same ontology an agent would produce structurally) ran end-to-end against actual RAI: telco, energy, supply_chain, machine_maintenance, portfolio all exit 0 with their headline outputs matching the runbook (telco $4,956,843 / OPTIMAL / 12 GOLD / 2 SILVER / 1 BRONZE, machine_maintenance OEE 79.8/68.2/61.4, supply_chain baseline $1,865, etc.).
• Live paste-test (supply_chain) — hand-wrote agent-style code from the trimmed prompt verbatim, ran against actual RAI. Concept set matches canonical (all 7). 5 of 7 concepts loaded with full row counts; Business (8/31) and Shipment (248/262) hit a NaN-row-dropping issue that the canonical script handles via batched loading. Not a trim regression — the original prompt didn't address this either; it's a runtime data-hygiene quirk an agent finds via 1-2 iteration cycles.

Conclusion: trims are safe. The agent generates equivalent loading code from either prompt form; the trim only removes redundant detail from the user's question.