feat(watch): allow custom agent commands to place the prompt explicitly #1707

Problem

squad watch --execute --agent-cmd is documented as a custom agent runner
override, but both watch command builders always append the Copilot-specific
-p <prompt> arguments.

That means a custom runner must imitate Copilot CLI's prompt contract. Runners
that use another prompt flag cannot be invoked directly. One motivating example
is Boundary, which accepts the task through --task; other task-oriented agent
CLIs have the same integration problem.

This also leaves a gap between:

• Add token usage tracking and cost visibility for agent sessions #453 / feat: add token usage tracking and cost visibility #454, which added token and cost visibility plus budget config; and
• Epic: Harden unattended watch execution boundaries #1477 / Spike: Verify sandbox and tool-governance coverage for unattended watch execution #1480, which call for enforceable unattended execution boundaries
  without inventing a Squad-specific policy language.

Proposed behavior

Allow one standalone {prompt} argv token inside a custom agentCmd:

boundary run ... --task {prompt}

When present, Squad replaces that token with the complete generated prompt as
one argv element. When absent, Squad preserves the current behavior and appends
-p <prompt>.

This is a generic command-contract fix. It does not make alternate engines
feature-equivalent to the default Copilot path.

Acceptance criteria

• agentCmd with no {prompt} behaves exactly as it does today.
• One standalone {prompt} token is replaced with the complete prompt as
  one argv element.
• Embedded forms such as --task={prompt} are not treated as placeholders.
• More than one {prompt} token fails before spawning with a clear error.
• Both buildAgentCommand and buildCopilotCommand use the same helper.
• Existing Windows argument escaping remains in the spawn path.
• Unit tests cover legacy, placeholder, duplicate-placeholder, and full
  multi-word prompt cases.
• Watch documentation includes one generic external-runner example and
  states that policy enforcement belongs to the runner.

Non-goals

• Adding Boundary as a Squad dependency.
• Implementing a second budget or sandbox policy engine in Squad.
• Claiming daily/weekly/monthly enforcement from a direct boundary run;
  Boundary currently exposes those windows through schedule/pipeline config.
• Preserving the default Copilot MCP injection for an alternate agent engine.
• Treating an estimated-dollar cap as exact GitHub Copilot billing.
• Replacing Squad's existing model category policy or cost reporting.

References

• Add token usage tracking and cost visibility for agent sessions #453
• Epic: Harden unattended watch execution boundaries #1477
• Spike: Verify sandbox and tool-governance coverage for unattended watch execution #1480
• packages/squad-cli/src/cli/commands/watch/agent-spawn.ts
• https://github.com/mavaali/boundary#spend-control