unic-spec-review S4: full review engine — all dimension + Green/Yellow/Red agents, Blue orchestrator, ranked hat-grouped triage

[orioltf]

Parent

Epic: #200 — unic-spec-review plugin.

What to build

Build out the full parallel review engine on top of the single-agent skeleton. Add the remaining dimension agents and the perspective agents, the Finding schema, and the ranking/grouping logic, then surface a ranked, hat-grouped triage.

• Dimension agents (Black-hat core): the seven not yet present — Ambiguity/Clarity, Spec-versus-Design, Spec-versus-Live, Internal-consistency, Testability, Feasibility/Constraints, Non-functional — alongside the existing Gaps/Completeness agent.
• Perspective agents: Green (alternatives the spec ignored), Yellow (stated value/justification), Red (likely user confusion). White is folded into Gaps + Testability.
• Blue orchestrator: fans out the agents in parallel and synthesises their Findings.
• Schema + ranking: finding defines the Finding schema (hat, dimension, confidence, severity, anchor target, body). finding-ranker sorts by confidence and severity. hat-mapper maps each dimension to its hat and groups Findings by hat for the report.

Every Finding is hat-tagged, dimension-tagged, Confidence-scored, and severity-scored. Output is presented conversationally as a ranked, hat-grouped triage and written to the markdown report (extending report-renderer to group by hat). The Landscape Brief from S3, where present, is injected into Testability, Feasibility, Spec-versus-Live, and Non-functional. Still read-only.

Note: Spec-versus-Design and Spec-versus-Live operate on whatever sources are available now (Confluence); Figma and live inputs arrive in S6.

Acceptance criteria

• All eight dimension agents plus Green/Yellow/Red perspective agents exist and are fanned out in parallel by the Blue orchestrator.
• finding defines the schema; every emitted Finding carries hat, dimension, Confidence Score, and severity.
• finding-ranker ranks Findings by confidence and severity; hat-mapper maps dimension→hat and groups Findings by hat.
• Triage output is presented conversationally ranked and hat-grouped; report-renderer writes the report grouped by hat.
• The Landscape Brief (when available from S3) is injected into Testability, Feasibility, Spec-versus-Live, and Non-functional agents.
• The run remains read-only.
• node:test unit tests cover finding-ranker, hat-mapper, the finding schema helpers, and report-renderer grouping with injected deps. The agent prompts and the Blue orchestrator are not unit-tested.
• PR adds CHANGELOG.md bullets under [Unreleased] and bumps the patch version; verify:changelog passes. No em dash in authored text (except the mandated CHANGELOG version header).

Blocked by

• unic-spec-review S1: skeleton tracer bullet — classify one URL, fetch page, one Gaps agent, write report #202 (S1 — single-agent skeleton, report renderer, classifier).