feat(skills): add context-intelligence-derived-metrics skill

[colombod]

Summary

Adds a new standalone Agent Skill, context-intelligence-derived-metrics, that teaches the graph-analyst to derive INTERPRETIVE metrics over otherwise-"blanket" graph/event data — data where many rows share one structural signature but carry different intent.

This skill documents the flavour metric (delegation transport × skill payload loaded inside spawned sub-session), joined via Delegation.sub_session_id = Session.node_id, and generalizes to magnitude-based and artifact-based discriminating joins. Applied to a real corpus, self-delegation dropped from ~42% to ~11.7% of delegations when one review-panel population was attributed out — proving the flat aggregate was an artifact; the derived dimension revealed the real story.

Sibling to context-intelligence-graph-query (which owns HOW to read the graph; this owns how to make blanket reads MEAN something).

Scope / guardrails

• agents/graph-analyst.md: intentionally UNMODIFIED — no wiring, soft pointers, or per-skill configuration needed. Auto-discovery via the tool-skills directory pointer is sufficient.
• Existing skills: no touch. Bundle integrity verified in DTU.
• Graph schema queries: depend only on documented properties (agent, skill_name, sub_session_id, FORKED/parent_id). Properties resolved_agent and SkillLoad.context are flagged as verify-before-cite (not part of this skill's queries).

Verification

• Skill auto-discovery proven — real DTU run: load_skill(list=true) inside the DTU lists context-intelligence-derived-metrics without any agent pointer, sourced purely via the tool-skills directory reference; the skill loads cleanly via load_skill(skill_name="context-intelligence-derived-metrics")
• No regression — all 8 existing sibling skills still list and load; discovery mechanism unchanged
• Seam crossed — tool/skill wiring (the auto-discovery boundary): verified in a Digital Twin Universe against the actual bundle and graph-analyst agent
• N/A — Module tests (modules/tool-context-intelligence-query doesn't apply; this is a skill, not a module that requires the client/server boundary tested)
• N/A — Top-level tests/ doesn't apply; this is a skill-methodology file with no executable tests (verification is query-outcome-based, conducted in the DTU and the earlier graph-analyst session)
• N/A — ruff check / ruff format not applicable; this is a pure-markdown SKILL.md file
• N/A — pyright not applicable; SKILL.md is markdown
• N/A — scripts/validate-full.sh bundle validation not required; this is a new file, no existing bundle structure changed

Real evidence on seams

Seam: Tool/Skill wiring (auto-discovery boundary)

Evidence: Digital Twin Universe run (amplifier-tester:setup-digital-twin).

• Snapshot pushed my working tree to isolated Gitea repo (admin/ci-derived-metrics-skill-test@main).
• Bundle's git source rewritten to point at that mirror.
• Inside DTU, ran: load_skill(list=true) → the new skill appears in the skill roster, discovered automatically.
• Ran: load_skill(skill_name="context-intelligence-derived-metrics") → returns the full body (243 lines), frontmatter parsed cleanly.
• Verification: On-disk grep of the mirrored graph-analyst.md inside the DTU → 0 references to the new skill (no soft pointer present).
• Regression: All 8 existing skills still list and load in the same DTU run.

Docs & diagrams

• No bundle.dot / bundle.png regeneration needed — bundle structure unchanged
• No README / skill contract changes — this is a NEW sibling skill, not an edit to an existing skill or the agent
• AGENTS.md / PR template / conventions not affected — this PR follows the established skill-addition pattern

Notes / follow-ups

• DTU ci-derived-metrics still running (awaiting cleanup).
• The skill is available for immediate use via load_skill(skill_name=...), no deployment lag — it auto-discovers as soon as it's merged to main and the bundle is re-sourced.