guidance: keep local development VMs small and necessary

[gaelic-ghost]

Problem

Current local-development guidance can lead an agent to create or enlarge a Linux VM/container runtime without first proving that virtualization is necessary or accounting for the developer machine's actual spare resources.

This caused an occasional Swift Lambda packaging task to use a Colima VM configured with a 100 GB virtual disk and 4 GiB RAM, even though ordinary TokenBrokerCloud development and integration testing can use native Homebrew PostgreSQL/Redis and does not need a VM. The stopped VM still occupied about 23 GB of host storage.

Requested guidance

Refine the applicable cloud deployment, server-side Swift, and general engineering guidance so agents:

• prefer already-installed native host services for local development and integration tests;
• do not introduce a VM merely to provide databases, caches, queues, or other services that already run natively;
• require a concrete OS/architecture/tooling boundary before using a VM;
• distinguish local runtime/testing from deployment-artifact packaging;
• make build/package VMs short-lived and on-demand;
• start from a conservative minimum (for example 1–2 CPUs, 2 GiB RAM, and a small explicitly sized sparse disk) based on measured workload needs;
• inspect and report both virtual disk capacity and actual host disk usage;
• ask before materially increasing CPU, RAM, or disk allocations;
• prefer hosted CI/build infrastructure for infrequent deployment artifacts when that avoids persistent local resource cost and is operationally reasonable;
• tear down disposable packaging environments when their cache value does not justify retained disk usage;
• never represent a VM resource maximum as free or harmless merely because allocation is sparse.

Acceptance criteria

• Relevant skills contain an explicit native-first decision gate.
• VM recommendations state a reasonable minimum and an escalation rule.
• Examples separate native service integration testing from Linux/architecture-specific packaging.
• Validation or review checks flag unjustified large defaults such as 4+ CPUs, large RAM reservations, or 100 GB disks.
• Guidance accounts for constrained developer laptops rather than assuming workstation/server capacity.