improvement(executor): run the Agent block on the shared model-fallback walk - #9982

Summary

• One shared fallback walk. Router, Evaluator and the Agent block now walk their model candidates through walkModelCandidates in executor/utils/model-fallbacks.ts. Before, the Agent kept its own copy of the walk. The walk owns:
  • the candidate order;
  • holding fallbacks until the final try;
  • de-duplication;
  • skipping unusable fallbacks (provider, permission, family);
  • resolving the saved credential once per call for candidates on the primary's provider;
  • the stop rules;
  • the block-log modelFallbacks record.
• Callers supply three pieces: an attempt, an optional prepare (a fallback that throws there is skipped, not counted as a failure), and skip/fallback log callbacks. executeModelRequestWithFallbacks is now a thin wrapper.
• The Agent keeps its own behavior in its callbacks:
  • Per-provider file hydration moves into prepare, which runs against the settled registries without mutating ctx. A fallback whose provider can't take the attached files is skipped.
  • Stream priming, the deep-research hold, the evaluation-model refusal and the log text are unchanged.
  • The manual save/restore of the error registries is gone. A fallback swaps ctx to the settled registries once, when its attempt starts.
• No behavior change for any of the three blocks.

Type of Change

• Improvement (refactor)

Testing

• The existing agent, router, evaluator and model-fallback-request tests pass unchanged; they serve as the behavior spec.
• New tests:
  • A primary failure followed by an attachment-skipped fallback rethrows the primary's error, redacted against the failed attempt's registries.
  • A later fallback runs from the settled registries.
  • The walk doesn't count a prepare skip as a failure, and rethrows the last real error when only skips remain.
• Red proofs:
  • Removing the registry swap or the skip semantics fails the new tests.
  • The new Agent tests also pass against the previous Agent loop, which shows behavior is preserved.
• Runs: 1,894 executor and lib/workflows/blocks tests pass. bun run lint, check-block-registry.ts origin/staging, check:test-patterns, check:comment-hygiene, check:explicit-any, check:boundaries, check:utils, check:file-names and check:tool-registry-boundary --check pass. Type-check and the full suite run in CI.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)