feat(widget): label tok/ctx% as substrate vs work (#32 tail)

[rz1989s]

What

The fleet widget's tok/ctx% segment is a context-occupancy metric dominated by the armory substrate (~59% from turn 1, barely moving across turns — see #32). On substrate-dominated runs it reads as "frozen / same every session," indistinguishable from a stuck run.

This adds a substrate vs work label to the widget's foreground run line: capture a substrateBaseline = contextTokens at the end of turn 1 (first assistant message_end), then once past turn 1 label the segment substrate when growth is ≤5% of the baseline (flat overhead) or work when tool results are growing the context.

▶ "review the PR"  · coder  5s  575K tok  59%  substrate  $0.0001
▶ "refactor module"  · coder  12s  655K tok  67%  work  $0.0042

Before turn 2 (or when no baseline was captured — e.g. turn 1 produced no assistant message) no label is shown: there's nothing to compare against yet.

Scope (intentional)

• Widget only (src/panel/widget-rows.ts) — enhancement: armory substrate consumes ~59% of subagent context window before work begins #32's pain is the live glance surface ("is it frozen or working?"). The Runs-tab (runs-rows.ts) is the navigable detail surface and is intentionally left unchanged; symmetric labeling there is a possible follow-up, not a blocker.
• fg runs only — bg runs don't carry substrateBaseline (no turnCount, different lifecycle) and never get the label.
• Live-only — substrateBaseline is a transient display heuristic, not journaled to the RunLog (same as feat: surface liveness progress for long foreground subagent runs #23's turnCount/lastEventClass).

Changes

• src/engine/run-registry.ts — RunRecord.substrateBaseline?: number (live-only, documented).
• src/engine/spawnSubagent.ts — capture once at the turn-1 first assistant message_end (turnIdx === 0); threaded into the runRegistry.update patch.
• src/panel/widget-rows.ts — SUBSTRATE_GROWTH_THRESHOLD = 0.05 + substrateBaseline on WidgetRun/toWidgetRun + the label in widgetLine (after ctx%, before cost).
• Tests — 6 widget cases (substrate flat / work growing / exactly-at-threshold / turn-1 no-label / no-baseline no-label / bg no-label) + 2 spawn cases (baseline captured at turn-1 message_end / not captured when no message_end).

Verification

• pnpm typecheck clean.
• pnpm test:run → 677/677 pass (669 existing + 8 new).
• Read-only review subagent verdict: Ship (no critical/important findings; 2 nits, both non-blocking and in-scope).

Closes part of #32 (direction 4 — widget UX). The other two #32 directions (memory-scope trim, prompt compaction) need a substrate measurement and ship as separate PRs.