token-usage: remaining accounting gaps after initial system

[Layau-code]

Summary

Token usage tracking system shipped in #2734 via #2770 #2800 #2838 #2841 #2845. This issue tracks remaining gaps found
during code audit.

Closes #2734

---

Bugs

1. model_name never persisted to RunRow

RunRow.model_name is used in aggregate_tokens_by_thread but never written — get_completion_data() omits it,
update_run_completion() doesn't set it. Result: by_model API always returns {"unknown": {...}}.

Fix: extract from serialized["kwargs"]["model"] in on_chat_model_start, return in get_completion_data().

2. SubagentTokenCollector returns early on multi-generation

token_collector.py:59 has a return after the first valid generation. Parent RunJournal.on_llm_end iterates all
messages. Behavior inconsistent.

Fix: remove the return, let dedup handle it.

3. llm_call_count excludes sub-agent calls

record_external_llm_usage_records adds to total_tokens but not _llm_call_count. Per-call average becomes
misleading.

Fix: increment count in record_external_llm_usage_records, or add separate counter.

---

UX

4. Sub-agent tokens invisible during streaming

Header shows accumulateUsage(messages) during active run. Sub-agent tokens aren't on any message's usage_metadata
— only visible after run completes and /token-usage is re-fetched.

Fix: emit periodic token usage SSE event during run.

---

Design

5. by_model granularity insufficient

Single run can use multiple models (lead/subagent/title). Current aggregation groups by one model_name per run.
Sub-agent model cost is invisible.

6. Sub-agent collector doesn't record model name

SubagentTokenCollector records have caller but no model_name. Combined with #5, per-model cost breakdown is
impossible.

---

Priority

P	Issue	Effort
0	#1 model_name	Small
1	#4 real-time display	Medium
1	#2 multi-gen fix	Small
2	#3 call count	Small
2	#5 #6 by_model redesign	Large

[WillemJiang]

@Layau-code, the PR #2775 addressed the model_name.