feat(ai): crew specialist agents — researcher/drafter/critic/editor (AI-041)

[mrviduus]

AI-041 — Crew specialist sub-agents (Phase 7: Multi-Agent Orchestration)

Four generic single-call `IAgent<TIn,TOut>` specialists that AI-040's `CrewOrchestrator` composes via `CrewTasks.Of`. Building blocks for AI-042 (AutoPublishCrew) / AI-043 (SeoCrew).

What

• `SingleCallAgent<TIn,TOut>` base — one `ILlmService.CompleteAsync` → one `llm_response` step → `AgentResult` w/ `AgentUsage(Iterations:1)`; `FeatureTag` (`crew.{role}`) flows to TracingDecorator.
• Researcher / Drafter / Critic / Editor — ~15 lines each, stateless singletons.
• Critic grounds on `ResearchNotes` to catch draft hallucinations; emits scored JSON (factual_accuracy/tone/length/banned_phrases + issues[]).
• `CriticOutputParser` is the keystone (AI-042 will gate publish on its blockers) — fail-closed: garbage/empty/unparseable → all-1s + a blocker issue + `ParseFailed:true`, never throws.
• Length + banned-phrases single-sourced via `BriefConstraints` so critic never flags what drafter wasn't told.

Why these decisions

• Deterministic parser over prompt-only trust — critic output gates a side-effect (publish), so the parser must degrade safely, not optimistically.
• Hardened per adversarial QA: string-aware balanced-brace extraction replaces naïve first-`{`..last-`}` (valid object + trailing prose, or `[{...},{...}]` array-wrap, no longer false-reject a good draft); unknown non-empty severity → `major` (not silently downgraded below the blocker gate); empty severity stays `minor` (absent ≠ asserted).

Tests — 43 added (full unit suite 356 green)

Per-agent output/usage/transcript mapping · CriticOutputParser adversarial battery (nested/in-string braces, escaped quotes, arrays, missing/null/float/NaN/string scores → worst-not-false-clean, severity coercion, never-throws) · 4-stage crew smoke through the real `CrewOrchestrator` (state threads research→draft→critique→edit, ordered transcript). StudyBuddy tool set-equality still green (no `ITool` leaked).

Verify

• `dotnet test tests/TextStack.UnitTests` → 356 pass
• `dotnet test tests/TextStack.AiEvals` → 24 pass / 5 skip (no-key floor)
• `dotnet format --verify-no-changes` → clean

🤖 Generated with Claude Code