Re-architect Step 7–9 agent capabilities

[Pierre Malarme (pmalarme)]

Summary

Re-architects the Step 7 multi-agent Coordinator/specialists and propagates the same capability split consistently through Steps 8 (workflow) and 9 (memory).

Capability changes

• Coordinator (Step 7) / finalize_itinerary step (Steps 8–9) now own both skills — the local travel-guide skill (renders the PDF trip guide) and the Foundry response-guardrails skill — and always use them.
• Hotels gains web search via the Foundry toolbox.
• Flights now always reports concrete fares/prices.
• Activities uses RAG + web only — the skills were removed.
• Guardrails are Coordinator/finalize-only (removed from the Activities specialist).

Consistency & housekeeping

• Kept the three surfaces in sync: docs → solutions → step_files/07 scaffold.
• Rewrote the 3 specialist prompt constants + their agent.yaml slices, and updated the agent.manifest.yaml files (specialist + root) to match the new ownership.
• Added agent-framework-orchestrations to step_files/00/requirements.txt.
• Added a Tip in the Step 7 doc to prefer full-size gpt-5.4 over -mini (context grows large with multi-agent handoffs).
• Reconciled mermaid diagrams, capability tables, and prose in the 07/08/09 docs and READMEs; carried the "skipped the Foundry skill?" fallback note into the Step 8 finalize section and added a sentence explaining why finalize is excluded from memory.

Validation

• python .workshop/scripts/lint_steps.py → 0 failures (2 pre-existing unrelated warnings).
• Custom parity check: every specialist prompt constant matches its slice verbatim; step_files/07 matches the Step 7 solution; all edited Python parses and all YAML loads.
• A full rubber-duck review found no blocking issues; its non-blocking findings (stale manifest/README wording) are addressed in this branch.

Co-authored-by: Copilot App 223556219+Copilot@users.noreply.github.com