Add a /research skill

[mxriverlynn]

Summary

This PR adds /research, a sizing-aware skill that takes an open-ended question and returns an evidence-backed, adversarially-validated report recommending an option without producing any committed artifact, so that Han has a question-shaped sibling to /investigate instead of overloading the bug pipeline.

• Adds the new /research skill (Han's 7th sizing-aware skill, 19th skill overall) plus a new research-analyst agent (22nd agent). It reuses codebase-explorer for codebase-grounded evidence and adversarial-validator to attack the result. The investigation behind this (recommendation.md) deliberately rejected expanding /investigate or building a two-mode skill, on Han's single-responsibility rule.
• Hardens the shared adversarial-validator agent with a new, generally-applicable 4th strategy ("challenge the evidence-gathering integrity"). This is the load-bearing change: it moves the web-reach threat model (indirect prompt injection, astroturfing, single-source laundering) from brief text into the agent's hardcoded contract, and also strengthens /investigate. Reviewers should weigh whether this 4th strategy is correctly scoped as additive and low-risk for existing consumers.
• Behavior controls worth scrutiny: web content is treated as data never instruction, the web-facing agent brief is isolated from codebase/operator context, evidence is required by default with an opt-in "exploratory" mode, and every report labels what does and does not have evidence.
• Mechanical fan-out across the repo: bidirectional "Does not X — use research" routing added to 5 neighbor skills (and their long-form docs), skill/agent counts bumped 18→19 / 21→22, sizing lists moved six→seven, and ~36 doc footers updated.

Behavior changes

• Before: an operator with an open-ended question ("what are my options for X", "should I use A or B", "how does Y work", "what's the prior art") had no skill that owned it. /investigate is a symptom→root-cause→fix pipeline; /plan-a-feature, /coding-standard, /gap-analysis, and /architectural-analysis each do research only as a bounded step toward a fixed artifact. There was no way to research options before committing to anything.
• After: /research <question> classifies and sizes the question, fans research agents out across the codebase, the open web, and provided material, consolidates a numbered evidence/artifact registry, builds an options landscape, recommends one option, runs an adversarial-validation pass that can overturn the recommendation, and writes a fixed-structure report (plain-language Summary → Research Results → Options → Recommendation+evidence-basis → Validation → Artifacts registry → References). It never emits a spec, standard, gap report, architecture assessment, or code. Out-of-scope, hybrid, and compound requests are routed or split rather than forced through. /investigate and 4 other neighbors now route research-shaped requests back to /research.

What to look at first

• The shared adversarial-validator 4th-strategy change (plugin/agents/adversarial-validator.md). This is the only edit that changes behavior of an existing, multi-consumer agent. CRIT-001 in the code review found the web-reach threat model was depending on brief text overriding the agent's closed 3-strategy contract; the fix promotes it into the agent. Confirm the "all three"/"minimum 5" wording was updated consistently and that it stays additive for /investigate.
• The web-reach trust model (plugin/skills/research/SKILL.md Operating Principles + Step 5, research-analyst.md anti-patterns). The skill reaches the live web, so untrusted content is a first-class input. Decisions D16 (data-not-instruction, context isolation, trust labeling) and D11 (corroboration, retrieval dates) are the defenses. WARN-001 specifically tightened Step 5's brief exclusion to match the stricter Operating Principle — worth confirming the two now agree.
• Evidence mode and report structure (D23, D24) — evidence required by default, opt-in "exploratory" mode, and one fixed report structure with inline artifact-ID cross-referencing. Reviewers should weigh whether the strict/exploratory split and the always-present Artifacts/References sections match the spec.
• Reviewer note (not a defect to fix here): the new files use em-dashes, which writing-voice.md bans unconditionally — but every existing plugin file already uses them. The code review surfaced this (WARN-004) as a repo-wide standard-vs-practice contradiction and deliberately did not de-em-dash /research in isolation, since that would make it the lone outlier. This wants a repo-wide decision, not a fix in this PR.

Files of interest

• plugin/skills/research/SKILL.md — the skill's behavior: classification, sizing, roster, web-reach controls, 8-step flow.
• plugin/agents/research-analyst.md — the new agent owning the web/prior-art and option-comparison angles, with the data-not-instruction anti-patterns.
• plugin/agents/adversarial-validator.md — the shared-agent 4th-strategy change; the only behavior edit affecting an existing consumer (/investigate).
• docs/plans/research-skill/recommendation.md — why /research is a separate skill, with the adversarial validation that corrected the original evidence.
• docs/plans/research-skill/artifacts/decision-log.md — D1–D24, where every behavioral tradeoff (web reach, roster, sizing, untrusted-source controls, evidence mode, report structure) was settled.