fix: attribute subagent volume to the model that dispatched it

[AqueGen]

What

tokens attributed every subagent to sessionModelOf(parent): the first model in the parent transcript's head, else the last in its tail. Long-lived sessions switch models with /model, so half of the agents were credited to a model the session was no longer on when it dispatched them, and the tokens report disagreed with the dispatch report (which stamps the model at dispatch time).

The session model of an agent is now resolved in this order:

1. The toolUseId from the agent's sidecar (agent-<id>.meta.json, already read for agentType) looked up in the parent transcript: the assistant line carrying that tool_use block is the dispatching message, and its message.model is the model in effect at that instant.
2. For nested agents (spawnDepth 2, parentAgentId in the sidecar) the same lookup in agent-<parentAgentId>.jsonl beside it.
3. The parent's assistant model with the greatest timestamp at or before the agent's first-line timestamp. Max timestamp, not last line in file order: a resumed session re-appends history, so file order is not chronological.
4. sessionModelOf as before.

One readFileSync per parent transcript, memoized by path; the walk reaches a subagents/ dir before the parent file beside it, so the index is built lazily. No hook change and no new dispatch-log field: the transcripts already hold everything needed, and this fixes history too.

The --session filter on main sessions is now per line model instead of per file, so a session that switched mid-way contributes only the volume that ran on the filtered model. Footer and commands/stats.md wording updated.

Why

Measured on this machine over 7 days: 269 agent transcripts, 269 sidecars, 269 with toolUseId; 264 resolve at step 1, the 5 remaining are nested agents. Attribution by the dispatching message vs by session head: fable-5-1 137 vs 17, opus-5 109 vs 232, sonnet-5 18 vs 8. The dispatch log over the same window: fable-5-1 138, opus-5 109, sonnet-5 22. So the two reports now describe the same population. Three independent re-measurements (Opus, Fable, Codex gpt-6) agreed; the timestamp rule and the toolUseId join disagreed on 0 of 264 agents, and no agent lacked a preceding parent model line.

Verification

node --test hooks/dispatch-counter.test.mjs hooks/update-check.test.mjs: 97 pass, 0 fail (93 baseline + 4 new: dispatch model over session head, nested agent via the parent agent's transcript, timestamp fallback, per-line --session on a main session). Existing fixtures carry no sidecars or timestamps and pass through step 4 unchanged.