feat(aidd-refine): adversarial and advisory agents

Start with plugins/aidd-refine/skills/02-challenge/SKILL.md, plugins/aidd-dev/agents/checker.md and executor.md, and docs/ARCHITECTURE.md. Trace the existing install/translate pipeline and its unit fixtures, then add the two agent sources and thin actions. Done means CI covers translation, reports enforce severity caps and floors, agents remain read-only, and README explains when to use each approach.

Problem

Refine exposes in-session recipes (challenge) but not isolated adversarial (attack assumptions, vacuous acceptance criteria, bypasses) and advisory (strengthen, prioritize) agents. Spec/plan stress-testing and constructive second opinions are improvised or skipped. The checker review path is a different job.

Scope

• Agents adversarial and advisory under aidd-refine/agents/ (canonical source format).
• Multi-tool presence = existing install/translate pipeline only (materialize into each installed tool’s native agent paths and formats). No symlinks across tool trees. Hosts never need to read .aidd/ for discovery.
• Thin skill/actions to run each agent on a named artifact (spec, plan, brief) with a fixed report shape.
• Severity floor + max findings (cap noise before any later specialist wiring).
• Read-only; never implement; never invoke orchestrator skills.
• Document when to use challenge vs adversarial vs advisory.

Standalone value: Spec/plan stress-test and second opinion without waiting for pluggable specialists.

Acceptance criteria

• Both agent source files ship in the plugin.
• Translator/unit fixture: canonical agent → at least one non-source target format (or identity when same format) — CI-able, not “UI discoverable”.
• Adversarial: findings with severity; no writes; respects cap/floor.
• Advisory: ordered recommendations; no writes; does not alone set a ship-blocking verdict.
• README: challenge vs adversarial vs advisory.
• Agents do not implement fixes or invoke orchestrator skills.

Prior art in this repo

• plugins/aidd-refine/skills/02-challenge/SKILL.md — in-session challenge recipe (not isolated agents)
• plugins/aidd-dev/agents/checker.md, plugins/aidd-dev/agents/executor.md — existing agent split (judge vs build)
• docs/ARCHITECTURE.md (~L157–160) — agents invoke recipes; cross-plugin by capability
• #548 — plan drift / challenge + acceptance-criteria gaps (user report)

Relations

Intended graph (same fields as GitHub parent / blocked-by / blocks). Formal GraphQL edges need maintainer write — see comment on #584.

Out of scope

• Auto-wiring into every review (pluggable specialists child)
• Project-local agents under .aidd/agents (materialize child)
• Symlink-based multi-tool layout