Composition notes

• Question: what is this one install — who runs it, on what, under which account — and what has it actually touched? The instance is the world's actor-web hub, so this is the hub page: every noun it touches has a row here that opens its Surface page.
• The merge: agent-detail-v2's divided header row (pawn + name · posture · spend · identity grid) over inventory-agent-expand-v2's organ set (overview → agent-scoped DataSurfacePanels → sessions), re-derived from the world. The expand card became the layer's AgentCard; this is the page it always pointed at.
• Verdict line is verdicts.agentDetail(world, id), one sentence, worst fact first; the second line is filled page-side from the posture items the greeting did not already name (agentDetailVerdict) — an install can carry an unreviewed server and a credential read and two off-list destinations, and the greeting only says one.
• Posture pill, not DataSmile (†): the comp's 0–100 Security Score has no derivation in the world, so a number feeding a smile would be invented. The pill's label counts facts — Standard (nothing), Elevated (attention items), Critical (a personal account or a redacted credential read) — and prints them beneath the label in plain text: the pill carries the tone, the items are its receipt. Vocabulary is open: Standard/Elevated/Critical is the c59 comp (qdash danger_level); ledger-v2 says Healthy/At Risk. Flagged here, not resolved (prior-art §7).
• Spend · 30d is estimated twice over and says so: the world carries measured 30d tokens per person, not per install, so the header apportions the person's total by this install's share of on-record session tokens and prints the method (sole / apportioned / unattributed / local / none). The 7-day receipt beneath is the measured number. If instance.tokens30d ever lands, the rule goes.
• Overview is seven claims, one sentence each, not a stat band — the numbers are small nouns (1 model, 3 servers), and seven blocks would fight the one-attention-device law. Labels stay grey; a row carries at most one amber fragment, the fact that needs attention. MCP prints Configured vs Connected with the undeclared / unreviewed / never-connected states in words.
• Panels are the six agent-scoped surfaces with rowTo into the Surface page (entityRoute(kind, id)) and allTo to the listing. Bars are this install's sessions on record referencing the row (models, servers, repos, destinations) or touches (files). Tools is the honest exception: the world knows which tools an install called, not how often, so its bars are fleet 30d calls and the caption says so — a T2 data ask. listingTo is unset: no listing page's protocol accepts an install as a filter, and the caps never trigger at this size.
• Honest states: unattributed renders as a first-class identity state (dashed "no identity", egress-only device), not a blank; Hermes says detected-only, no account, local $0, and its Ollama pulls still show under Endpoints because the device is sensored; Wes's never-run install prints seven empty sentences and "never run — nothing to estimate". An id the current world lacks (Dana's personal ChatGPT under all-clear) gets the verdict's "no such install" and the switcher.
• Derivation: every number traces to derived.ts or derived/agent-instance.ts (agentSpend / agentPosture / agentSurface / agentSessionRows); nothing hand-typed. Links carry ?state=; links into pages not yet built are expected under the c81 protocol.