fix(dry-run): return schema-valid mock responses to prevent grader crashes

[christso]

Summary

• VSCode provider dry-run mode returned output: [], which produced an empty candidate string that caused all built-in graders to fail or report errors
• Fixed both invoke() and invokeBatch() to return output: [{ role: 'assistant', content: '{}' }] plus zeroed tokenUsage: { input: 0, output: 0 }
• Updated AGENTS.md to reflect that dry-run now returns schema-valid responses

Before / After

Red (before fix): Dry-run response has output: [], so extractLastAssistantContent returns '':

• is-json grader: fails — '' is not valid JSON
• execution-metrics with max_tokens: fails — "Token usage data not available"
• contains/equals/regex: trivially fail on empty string

Green (after fix): Dry-run returns output: [{ role: 'assistant', content: '{}' }] with tokenUsage: { input: 0, output: 0 }:

• is-json grader: passes — '{}' is valid JSON
• execution-metrics: has zeroed token data, so it evaluates normally
• contains/equals/regex: operate on a non-empty string, no crash

Test plan

• 4 regression tests added in vscode-provider-dry-run.test.ts covering invoke() and invokeBatch() for all three properties: non-empty output, valid JSON content, zeroed tokenUsage
• @agentv/core test suite: 1635 pass, 0 fail
• Lint passes

Closes #1088

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	cdcee62
Status:	✅  Deploy successful!
Preview URL:	https://3438a616.agentv.pages.dev
Branch Preview URL:	https://fix-1088-dry-run-mock-schema.agentv.pages.dev

View logs

[christso]

Red/Green UAT Evidence

Red (without fix): Stashed vscode-provider.ts fix and ran regression tests:
```
(fail) VSCodeProvider dry-run response shape > returns non-empty output so graders do not crash
(fail) VSCodeProvider dry-run response shape > returns valid JSON content so is-json grader passes
(fail) VSCodeProvider dry-run response shape > returns zeroed tokenUsage so execution-metrics grader does not report missing data
(fail) VSCodeProvider dry-run response shape > batch invoke returns the same schema-valid shape per response
1631 pass / 4 fail
```

The failures confirm the bug: output: [] → empty candidate → is-json fails, tokenUsage undefined.

Green (with fix): Applied vscode-provider.ts changes:
```
1635 pass / 0 fail
```

All 4 regression tests pass. The dry-run response now returns {} (valid JSON) with tokenUsage: { input: 0, output: 0 }.