fix: revert .optional() to .default([]) and strip empty evo arrays on write

[notgitika]

Summary

Fixes the JSON Schema generation failure introduced by #1073. The .optional() change causes Zod v4's toJsonSchema() to throw "Undefined cannot be represented in JSON Schema" during the release workflow.

Approach

• Schema: Revert configBundles, abTests, httpGateways back to .default([]) so JSON Schema generation and Zod parsing work correctly
• writeProjectSpec: Strip empty arrays for these three fields before writing to disk, so preview features don't appear in agentcore.json for users who haven't opted in

This gives us both: correct schema generation AND clean user files.

Test plan

• npm run typecheck passes
• Pre-commit hooks pass (eslint, prettier, secretlint)
• npm run generate-schema succeeds (no "Undefined cannot be represented" error)
• agentcore create does not write configBundles/abTests/httpGateways
• agentcore add config-bundle correctly writes configBundles

[agentcore-cli-automation]

The empty-array stripping in writeProjectSpec won't actually keep these fields out of agentcore.json. The stripping runs on the input data, but validateAndWrite then calls schema.safeParse(data) and writes result.data — and because the schema uses .default([]), Zod re-populates the stripped fields with empty arrays on parse. The written JSON will still contain "configBundles": [], "abTests": [], "httpGateways": [].

Quick trace with the agentcore create path:

1. createDefaultProjectSpec() returns an object without configBundles/abTests/httpGateways (see src/cli/project.ts).
2. writeProjectSpec(spec) — the three cleaned.xxx?.length === 0 checks are all false (the fields are undefined, not empty arrays), so nothing is deleted.
3. validateAndWrite → AgentCoreProjectSpecSchema.safeParse(cleaned) applies .default([]) for each of the three fields.
4. JSON.stringify(result.data, null, 2) writes configBundles: [], abTests: [], httpGateways: [] to disk.

So the second objective in the PR description ("writeProjectSpec: Strip empty arrays... so preview features don't appear in agentcore.json for users who haven't opted in") isn't met as written.

A few options to actually achieve that:

1. Strip after validation. Parse with the schema, then delete the empty arrays from result.data before stringifying. This would likely require splitting the logic out of the generic validateAndWrite (e.g., do a manual parse + post-process + write in writeProjectSpec), or adding an optional post-validate hook.

2. Use a separate write schema that keeps the fields .optional() (no default) for serialization, while the read/runtime schema keeps .default([]). The schema used by generate-schema.mjs (the one that was breaking) can stay .default([]); the write path uses the optional variant.

3. Keep .optional() everywhere (the #1073 approach) and fix the JSON-Schema generation failure differently. Since generate-schema.mjs already strips defaults from required, the real blocker was z.toJSONSchema throwing on undefined. That can typically be handled by wrapping the optional arrays with .default([]) only at JSON-schema-generation time, or by pre-processing the schema before calling z.toJSONSchema.

Option 1 is probably the smallest change. Either way, the current PR doesn't produce the intended agentcore.json output — worth adding a test that reads the file after agentcore create and asserts these keys are absent, which would catch this.

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	42.89%	8936 / 20830
🔵	Statements	42.17%	9485 / 22491
🔵	Functions	39.66%	1537 / 3875
🔵	Branches	39.82%	5769 / 14486

Generated in workflow #2257 for commit 5f1c177 by the Vitest Coverage Report Action