agents(skills): add issue-triage skill for maintainer issue classification and hygiene

Problem

Deliberate Lab receives bug reports, feature requests, and meta-tooling proposals from a diverse set of contributors, researchers, and maintainers. Across the backlog of open issues:

• Over 80% lack Conventional Commit prefixes or scopes, using unstructured phrases (e.g. "Agents will skip the stage without answering questions").
• Issues frequently lack ADR 0004 layer labels (runtime:*, area:*), or carry legacy catch-all labels like process.
• Issue titles authored by AI coding agents frequently suffer from prefix drift (e.g. using feat(skills) or fix(build) rather than agents(skills) or build(lint)).
• Descriptions often omit key context needed for implementation (e.g. whether a bug touches participant UX vs. experimenter UX, reproduction steps, or edge-case expectations).

Currently, maintainers must manually read each issue, mentally map it against repository taxonomy, construct conventional titles, and apply labels by hand via separate terminal commands.

Proposed Solution

Introduce a dedicated issue-triage pairing skill (.agents/skills/issue-triage/SKILL.md) that equips an AI agent to inspect and triage any GitHub issue on command.

1. Invocation

Trigger phrases:

• "Let's triage issue "
• "Triage issue "
• "Audit issue "

2. The 4-Point Triage Rubric

Upon invocation, the agent inspects the issue losslessly using gh-issue-view.sh (per ADR 0003) and evaluates it against four standardized criteria:

1. Architectural Layer Classification (ADR 0004):
   • Identifies which of the 6 repository layers the issue touches:
     • Product/Runtime: runtime:participant-ux, runtime:experimenter-ux, runtime:agents, runtime:backend
     • Developer/Operational: area:workspace, area:git, area:build, area:test, area:ci-deploy
2. Conventional Title & Prefix Hygiene (#1258):
   • Transforms unstructured or misnamed titles into canonical type(scope): description.
   • Enforces runtime gating: feat: and fix: reserved strictly for user-facing product runtime.
   • Enforces developer tooling types: agents(...) for skills/workspace/decisions, build(...) for lint/build, chore(...) for repo-wide governance.
3. Label Applicability & Audit:
   • Recommends labels to add (e.g. specific area:* or runtime:* tags).
   • Recommends legacy or conflicting labels to remove (e.g. retiring generic process).
4. Clarity, Completeness & Repro Audit:
   • Evaluates whether problem statement, reproduction steps, expected behavior, and scope are clear.
   • Formulates targeted clarifying questions for the author if critical details or edge cases are underspecified.

3. Mandatory Human Authorization Gate

• The skill presents a compact, formatted triage summary with the proposed title and label diff.
• The agent strictly waits for explicit maintainer approval before executing any mutations (gh issue edit, gh issue comment).
• When approved, updates are applied in a single shot.

Related Context

• ADR 0003 (0003-agent-tooling-output-standards.md)
• ADR 0004 (0004-repository-layer-taxonomy.md)
• #1247 (Maintainer PR evaluation protocol)
• #1258 (Adoption of agents(...) Conventional Commit prefix)

Create .agents/skills/issue-triage/SKILL.md and first read ADR 0003, ADR 0004, #1247, and #1258. Use the stated gh-issue-view.sh workflow and document the four-point rubric, invocation phrases, and human authorization gate before gh issue edit or gh issue comment. Done means the skill produces a compact title and label diff, asks clarification questions when needed, and waits for explicit approval before mutations.