Require explicit Copilot engine for init artifacts - #50427

pkg/cli/init_command_test.go:592

• The non-Copilot regression now expects the dispatcher skill to exist, contrary to the requirement that non-Copilot initialization remain free of Copilot-specific artifacts. Assert non-existence instead.

	if _, err := os.Stat(filepath.Join(".github", "skills", "agentic-workflows", "SKILL.md")); err != nil {
		t.Errorf("Expected dispatcher skill file to be created for non-Copilot engine: %v", err)

Key Themes

• Shared test directory causes temporal coupling: The new two-phase test runs both default and --engine copilot init in the same tmpDir. The second phase can't falsify the first phase's absence assertions because those files may already be present.
• Stale MCP: true / Agent: true in non-Copilot test calls: A dozen existing tests still pass those options without Engine: "copilot", but don't assert that the corresponding artifacts are absent — silent regression risk.
• Dropped log line: The old "Skipping Copilot dispatcher skill" observability is now gone.

Positive Highlights

• ✅ Core logic change is a clean one-liner: opts.Engine == "copilot"
• ✅ Help text updated accurately — the Copilot-specific flags are now clearly documented under --engine copilot
• ✅ Good coverage added for the new default-engine behaviour
• ✅ Deprecated --mcp flag message updated to guide users to --engine copilot

[/tdd] Several tests (TestInitRepositoryErrorHandling, TestInitRepositoryWithExistingFiles, TestInitRepositoryWithCodespace, etc.) still call InitRepository without Engine: "copilot" but include MCP: true and Agent: true. Under the new semantics those Copilot artifacts won't be created, yet no assertion confirms they're absent — a future regression could re-enable them silently.

<details>
<summary>💡 Suggestion</summary>

Add a negative assertion in each non-Copilot call site…

[/tdd] The new two-phase test (default run, then --engine copilot run) shares a single tmpDir directory. This means the second run inherits all files written by the first. It cannot detect whether the Copilot artifacts were absent before --engine copilot was passed, because the first run's output is already on disk.

<details>
<summary>💡 Suggestion</summary>

Split into two independent subtests, each with its own TempDir:

t.Run("default engine", func(t *testing.T) {
    //…

</details>

<details><summary>pkg/cli/init.go:10</summary>

**[/codebase-design]** The dispatcher skill is now unconditionally created regardless of engine (moved outside the `copilotArtifactsEnabled` block), but the log message on line 31 still says &quot;Writing agentic workflows dispatcher skill&quot; without mentioning why it runs for all engines. The previous else-branch logged `&quot;Skipping Copilot dispatcher skill for engine: %s&quot;` — that message has been dropped entirely.

&lt;details&gt;
&lt;summary&gt;💡 Suggestion&lt;/summary&gt;

Add a brief log message for the non-Copilot…

</details>

In pkg/cli/init.go line 132:

if opts.MCP && copilotArtifactsEnabled {

copilotArtifactsEnabled is opts.Engine == "copilot" (line 45). Previously (opts.Engine == "" || opts.Engine == "copilot") the default/no-engine case still allowed MCP artifacts. Now, gh aw init --mcp alone produces no .github/mcp.json or copilot-setup-steps.yml, silently diverging from the flag's documented "backward compatibility" purpose. There's no validation error or warning when --mcp is set without --engine copilot, so users relying on scripted/CI usage of the deprecated flag get silently different behavior after upgrading.

Suggested fix: either emit an explicit error/warning when --mcp/true is set without --engine copilot, or treat explicit --mcp as also implying Copilot artifact generation for true backward compatibility.

Deprecated --mcp flag now silently no-ops for the default engine, breaking backward compatibility.