Review pipeline: calibrate review frequency to step risk level

[sansari]

Problem

The DeepWork review pipeline applies uniform review intensity across all workflow steps regardless of their risk profile. This causes significant wall-time overhead on small/personal analyses where the review cost far exceeds the value — while still being appropriate for complex, high-stakes deliverables.

Identified from session analysis of a personal-site analysis project (salmanio-analysis). Three distinct categories of wasteful review were identified:

• Category A — Redundant validation: DocBaker already validates its own output via validation_report.json, yet a separate deepwork:reviewer agent then re-reads the same screenshots. ~8 minutes wasted per DocBaker report.
• Category B — Trivial step reviews: Config files, YAML decisions files, and env file generation steps get full reviewer agents despite being trivially reversible and low-risk. 22s–2m10s per occurrence.
• Category C — Failed reviewers with no hard timeout: Reviewer agents that hit API issues return 0 tool uses after ~4 minutes each, wasting wall time and producing nothing.

Proposed approach

Proportional review: match review depth to actual step risk rather than applying the same intensity everywhere. Full reviews should be preserved on final deliverables and high-stakes analysis outputs. See sub-issues for each category.

Sub-issues

• Category A: Category A: Skip present_report reviewer when DocBaker validation_report.json already passed #406
• Category B: Category B: Lightweight validation for trivial/reversible workflow steps #407
• Category C: Category C: Hard timeout and retry for failed reviewer agents #408