docs(claude-md): trim CLAUDE.md duplication to invariants (−346 always-tok)

CLAUDE.md is loaded every turn while working in this repo (measured 2,425
always-loaded tokens — the whole per-repo delta; no .claude/rules or .mcp.json).
The file is already dense (mostly tables), so the trim targets duplication and
impl-detail prose, not domain content:

- KB-routing: dropped the per-agent core-file lists — line 81 already states the
  load contract is "definert i agentfilen", so the per-agent enumeration
  duplicated the agent files (single source of truth). Kept the principle + the
  ros-analysis-agent exception + the summary-agent note.
- Recommended MCP servers: collapsed to one line + the existing reference-doc
  pointer (the detail lives in references/architecture/recommended-mcp-servers.md).
- Spor-D hook note: 7 lines → 1 (impl detail about the cached score report).
- Future sibling-plugins: bullet list → one line (speculative roadmap).

Kept intact: command (29) / agent (12) / skill (5) tables, the ## Hooks section
+ both hook script names (tests/test-plugin-discovery.sh asserts these), and the
EU AI Act deadline table + Digital Omnibus note (maintained domain data, possibly
deliberate always-loaded — left for the operator to decide if it moves to a
reference doc; ~350 tok further trim available there).

Verified: tests/test-plugin-discovery.sh 13/13, tests/validate-plugin.sh 239/0
("VALIDATION PASSED"). CLAUDE.md 147→126 lines, 9,843→8,444 B, 2,425→2,079 tok
(−14%). Docs-only — no version bump, no catalog ref change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01683eAqVecv9VZfQzL8CQ9h

CLAUDE.md

@@ -78,14 +78,7 @@ Tilbyr strukturert arkitekturveiledning for Microsoft AI-stakken:
 
 ### Kunnskapsbase-routing i agenter (typisk 3 kjernefiler per invokasjon)
 
-Agenter leser navngitte kjernefiler, ikke hele kataloger. «3 kjernefiler» er normalen for review-/security-/cost-mønsteret; **ros-analysis-agent er et dokumentert unntak** med et større, eksplisitt last-kontrakt (kjerne + betinget) definert i agentfilen — fordi en deterministisk ROS krever trusselbibliotek + rubrikker + metodikk + mal som minimumskjerne.
-- **security-assessment-agent**: security-scoring-rubrics-6x5.md, ai-security-scoring-framework.md, ai-threat-modeling-stride.md
-- **cost-estimation-agent**: deterministic-cost-calculation-model.md, azure-ai-foundry-cost-governance.md, cost-models.md
-- **architecture-review-agent**: decision-trees.md, security.md, public-sector-checklist.md + domene-spesifikke ved behov (RAG/MLOps lastes betinget)
-- **ros-analysis-agent**: kjerne (alltid, 4): ros-ai-threat-library.md, ros-scoring-rubrics-7x5.md, ros-methodology-ns5814-iso31000.md, ros-report-templates.md + betinget på trigger (sektor / MAESTRO / DPIA-integrasjon / AI Act) — se last-kontrakt i agentfilen
-- **dpia-agent**: dpia-norwegian-methodology-ai.md, gdpr-compliance-ai-systems.md, ai-impact-assessment-framework.md + betinget cross-border (Schrems II → data-residency-audit-monitoring.md med EDPB seks-stegs-TIA)
-- **ai-act-assessor**: ai-act-classification-methodology.md + relevante ai-act-*.md filer (maks 3 per fase)
-- **summary-agent**: Leser assessment-outputs fra sesjon, ikke KB-filer
+Agenter leser navngitte kjernefiler, ikke hele kataloger. «3 kjernefiler» er normalen for review-/security-/cost-mønsteret; **ros-analysis-agent er et dokumentert unntak** med et større, eksplisitt last-kontrakt (kjerne + betinget) — en deterministisk ROS krever trusselbibliotek + rubrikker + metodikk + mal som minimumskjerne. Hver agents nøyaktige kjerne-/betinget-filer er definert i **agentfilen** (single source of truth); `summary-agent` leser sesjonens assessment-outputs, ikke KB-filer.
 
 ## MCP-servere
 
@@ -94,11 +87,7 @@ Agenter leser navngitte kjernefiler, ikke hele kataloger. «3 kjernefiler» er n
 
 ### Anbefalte MCP-servere (ikke påkrevd)
 
-- `azure-mcp-server` (microsoft/azure-mcp-server) — Live Azure-infrastrukturinspeksjon (Storage, Key Vault, Monitor, AI Search, RBAC)
-- `bicep-mcp-server` — IaC-generering for Azure-ressurser
-- `azure-devops-mcp` (microsoft/azure-devops-mcp) — Work items, pipelines, repos
-
-Se `references/architecture/recommended-mcp-servers.md` for detaljer.
+`azure-mcp-server` (live Azure-infra-inspeksjon), `bicep-mcp-server` (IaC), `azure-devops-mcp` (work items/pipelines/repos). Detaljer: `references/architecture/recommended-mcp-servers.md`.
 
 ## Hooks (2)
 
@@ -107,13 +96,7 @@ Se `references/architecture/recommended-mcp-servers.md` for detaljer.
 | SessionStart | `session-start-context.mjs` | Viser aktive utredninger, KB-ferskhet, onboarding-status, skill- + kurs-signaler (Spor C), skill-kvalitetssignaler (Spor D — skills under 90 %, fra cachet score-rapport) + AI Act-frister |
 | Stop | `stop-assessment-reminder.mjs` | Påminnelse om ucommittede vurderinger, neste steg |
 
-> Skill-kvalitetssignalet (Spor D) leser den **cachede** rapporten
-> `scripts/kb-eval/data/skill-score-report.json` og scorer aldri live
-> (buildReport leser ~400 ref-filer — for dyrt i en SessionStart-hook).
-> Cachen er gitignored (derivert/regenererbar via `score-skill.mjs --write`),
-> så den er fraværende i en fersk klon: linja surfacer kun for en vedlikeholder
-> som har produsert cachen lokalt, aldri for en sluttbruker. Alle skills ≥90 %
-> ⇒ ingen linje.
+> Spor D-signalet leser den **cachede** `scripts/kb-eval/data/skill-score-report.json` (aldri live — buildReport leser ~400 ref-filer); gitignored/regenererbar via `score-skill.mjs --write`, så fraværende i fersk klon → kun synlig for en vedlikeholder, og kun for skills <90 %.
 
 > Secrets scanning consolidated to llm-security plugin.
 
@@ -140,8 +123,4 @@ Se `references/architecture/recommended-mcp-servers.md` for detaljer.
 
 ## Relaterte plugins (fremtidig)
 
-- `ms-rag-architect` — RAG-spesialist (egen plugin)
-- `ms-power-automate-architect` — Power Automate deep-dive
-- `ms-azure-ai-architect` — Azure AI Services deep-dive
-- `ms-foundry-architect` — Microsoft Foundry spesialist
-- `ms-copilot-studio-architect` — Copilot Studio spesialist
+Planlagte søsken-plugins: `ms-rag-architect` (RAG), `ms-power-automate-architect`, `ms-azure-ai-architect`, `ms-foundry-architect`, `ms-copilot-studio-architect`.