fix(agent): bound generate_scene retries per page

[hichipli]

Summary

Bound repeated generate_scene failures for the same page within one server-workbench run. The initial generation and one retry may execute; further consecutive calls for the same stageId + order are rejected before reaching the generation provider.

Related Issues

Fixes #1308.

Changes

• add a request-scoped retry budget keyed by the semantic page target (stageId + order)
• preserve the original failure content/details while adding structured retry state and clear recovery guidance
• block calls after the second consecutive failure, before another generation request can run
• reset the budget after a successful generation or when generation moves to another page
• keep unrelated tool calls outside the budget and keep the shared tool allowlist authoritative
• add focused policy, Agent-hook, and runner-wiring regression coverage

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update
• Refactoring (no functional changes)
• CI/CD or build changes

Verification

Steps to reproduce / test

1. Run the focused retry-policy, Agent-hook, and runner-wiring tests.
2. Run the complete agent-runtime test directories under Node 22.
3. Run the complete repository test suite under Node 22.
4. Run formatting, lint, and TypeScript checks.

What you personally verified

• the first failed page generation leaves exactly one real retry available
• a second consecutive failure returns exhausted-state guidance and a third call is blocked before tool execution
• a successful generation resets the budget
• moving to a different page target resets the budget
• unrelated tools neither consume nor reset the active page budget
• existing error details are preserved when retry metadata is added
• the shared allowlist still runs before the new request-scoped pre-call hook
• Node 22 full suite: 652 test files passed, 9 skipped; 7,170 tests passed, 81 skipped
• Node 22 agent-runtime suites: 85 test files passed, 4 skipped; 1,044 tests passed, 11 skipped
• local lint completed with 0 errors; 18 unrelated baseline warnings remain

Evidence

• Before: every new tool-call ID for the same failing page could reach generation again.
• After: only the initial attempt and one retry can execute consecutively for a page; later same-target calls return a deterministic blocked result while later page targets remain available.
• CI passes (pnpm check && pnpm lint && npx tsc --noEmit)
• Manually verified locally with focused and complete automated suites
• Screenshots / recordings attached (no UI changes)

Checklist

• My code follows the project's coding style
• I have performed a self-review of my code
• I have added/updated documentation as needed
• My changes do not introduce new warnings