feat: trace cron-fired agent runs alongside chat

[nshi]

Summary

Cron-triggered agent jobs were silently invoking record_event without a scoped trace context, so events landed in runtime-trace.jsonl with session_id=None and were orphaned in the viewer (`dev/trace-viewer.py`). This change makes cron runs first-class in the trace pipeline — same detail and granularity as chat — and clearly distinguishes them in the viewer.

• Unified trace context: replaced `RUNTIME_TRACE_SESSION_ID: Option` task-local with a typed `Option<Arc>` carrying `kind` (Chat | Cron), `session_id`, and cron job id/name. Both the channels boundary and the cron scheduler now wrap the agent loop in the same scope; the agent loop itself is unchanged. `Arc` keeps `record_event` cheap at the ~26 call sites in the agent loop (pointer clone, not 4× String clone).
• Cron lifecycle bookends: `cron_job_started` (with job_id/name/schedule/prompt/model_override/run_id) and `cron_job_finished` (with success/duration_ms/output_preview) frame each firing — analogue of `channel_message_inbound/outbound` for chat. Heavy payload construction is gated behind `runtime_trace::is_enabled()` so it's skipped when `runtime_trace_mode = "none"`.
• Viewer support: `Turn` and `ConversationSession` track `kind` / `cron_job_id` / `cron_job_name`. Cron sessions render with a yellow `[CRON]` tag and `job=` instead of channel. New `--kind {chat,cron,all}` filter. `search_sessions` now matches cron metadata as well as inbound chat content.
• Backwards compatible: legacy JSONL events without `kind` parse via serde defaults (default = Chat).

Test plan

• `cargo fmt --check` — clean
• `cargo clippy --all-targets -- -D warnings` — clean
• `cargo test --lib` — 4281 tests pass; new tests cover (a) legacy events without `kind` deserialize as Chat, (b) `record_event` under `TraceContext::cron` populates kind/job_id/job_name and `session_id == run_id`
• Viewer smoke test on a synthesized JSONL: `--kind cron`, `--kind chat`, and `search ` all render correctly with the `[CRON]` tag
• Manual end-to-end (after merge): schedule a one-shot agent cron `mentat cron add agent --at "+30s" --prompt "..."`, wait for it to fire, then `tail ~/.mentat/workspace/state/runtime-trace.jsonl | jq 'select(.kind=="cron")'` and confirm `cron_job_started` → `provider_api_request` → `cron_job_finished` all share the same `session_id`. Then `uv run dev/trace-viewer.py recent --count 5` should show the run with the `[CRON]` tag.