fix(tui): include sub-agent costs in session cost

[jkyberneees]

What

Sub-agent LLM spend was missing from every session-cost surface: the header, the /stats cost row, and the cockpit cost-cap row all computed costUSD(sessCtxTok, sessOutTok, prices) — main-loop tokens only. Each finished sub-agent's engine-reported cost_usd (wire v2 P6) decorated its card in the transcript but never reached the session figure.

Fix

• Bank each finished task's final cost per task_id (Model.subCosts), upserted from finished subagent_state frames — replayed frames overwrite instead of double-counting.
• Banking happens before the card-attach attempt, so stray/resumed/late frames still count — spent is spent.
• Add subCostTotal() to all three surfaces (header, /stats, cockpit cap row — the cap judges total spend).
• Reset with the other session telemetry on session swap (resume) and /clear — no cross-session leakage.
• Absent/zero cost records nothing: unavailable is never $0.

The framed result envelope stays display-only (it carries no task_id, so it cannot upsert safely); state frames are the single source of truth.

Tests (RED-first)

6 new tests in internal/tui/subagent_cost_test.go; the 4 behavioral tests were confirmed failing before the fix:

• TestSubagentCostAddsToHeader — header = main + sub ($0.016 + $0.0175 = $0.0335)
• TestSubagentCostReplayIsIdempotent — replayed finish frames never double-count
• TestStatsCardIncludesSubagentCost — /stats cost row carries sub-agent spend
• TestCockpitCapRowIncludesSubagentCost — cap row compares total spend against the cap
• TestSubagentNoCostReportsNothing — guard: absent cost never distorts the figure
• TestClearResetsSubagentCost — guard: /clear wipes banked costs

Verification

make fmt · make vet clean · make lint 0 issues · make test (go test -race ./...) all green.