Migrate eligible model stages to structured output

[eshulman2]

Problem

Several Forge stages request a specific machine-readable structure in their prompts and then parse free-form model text. This is fragile: formatting deviations cause avoidable parsing failures, retries, and unclear errors.

Forge currently does not use provider-enforced structured output. The model-policy capabilities field therefore correctly supports only tools; structured_output must not be introduced until Forge actively uses and validates it.

Goal

Use schema-enforced structured output for stages whose results are genuinely data-shaped, while retaining Markdown/text output for narrative artifacts.

Candidate stages

Evaluate at least:

• Bug, automated-review, proposal-review, and task-takeover triage
• Epic decomposition
• Task generation
• CI analysis decisions
• Review-thread classification and other bounded decision outputs

Keep narrative stages such as PRDs, specifications, implementation plans, PR descriptions, and qualitative reviews as text/Markdown unless a clear schema boundary is identified.

Requirements

• Define typed Pydantic response schemas for each migrated stage.
• Apply structured output to the final response without disrupting the Deep Agent tool loop.
• Verify behavior across every supported backend: Vertex AI Gemini, Vertex AI Anthropic, Google GenAI, and direct Anthropic.
• Preserve useful retries and a clearly documented fallback for provider/schema failures.
• Replace ad hoc text parsing only after equivalent behavior and error handling are covered by tests.
• Produce actionable validation errors rather than silently accepting malformed output.
• Keep Langfuse model and stage attribution consistent.
• Introduce a structured_output model-connection capability only when runtime enforcement exists and a stage concretely requires it.
• Document backend limitations and configuration requirements.

Acceptance criteria

• Eligible stages return and consume validated typed objects.
• Tool-enabled stages can still complete their tool loop before producing the final structured response.
• Tests cover valid output, malformed output, retry/fallback behavior, and all supported provider adapters.
• Existing workflow behavior remains backward compatible.
• Remaining prompt-parsed stages are documented with the reason they were not migrated.