fix: serve/web runs never persist turn metrics or context estimates

[sam-saffron-jarvis]

What changed

• Updated cmd/serve_runtime.go so the existing SetTurnCompletedCallback now persists per-turn metrics for serve/web runs via store.UpdateMetrics(...).
• Persisted the engine context-estimate baseline on normal turn completion via store.UpdateContextEstimate(...), matching the existing CLI/TUI/Telegram behavior.
• Updated the in-memory sessionMeta counters and persisted context estimate when those writes succeed so active serve sessions stay consistent with stored session state.
• Added a regression test in cmd/serve_runtime_test.go that exercises a multi-turn tool run and verifies both metrics and context estimates are persisted.

Why this is high-value

Serve/web sessions were successfully appending transcript rows but silently leaving session accounting stale on normal completion. That meant:

• llm_turns, tool_calls, and token totals could stay zero or outdated
• last_total_tokens / last_message_count were not saved for resumed web sessions
• user-visible stats and any usage/cost reporting based on session rows became incorrect
• resumed sessions could start with a bad context estimate even though other surfaces already persisted it correctly

This fix is small, targeted, and aligns serve/web with already-shipped persistence behavior in ask/TUI/Telegram.

Validation

• Added regression test: go test ./cmd -run TestServeRuntimeRunPersistsTurnMetricsAndContextEstimate -count=1
• Ran gofmt -w cmd/serve_runtime.go cmd/serve_runtime_test.go
• Ran go build ./...
• Ran go test ./...
• Ran git diff --stat and git diff --check