Every agent call names the model that ran it

[czpython]

The sibling of #144. AgentCall.model was nullable on stated grounds:

# Nullable — a call may be recorded before its model is resolved.
model: Mapped[str | None] = mapped_column(String, default=None)

It can't be. _run resolves model = self.get_model_name() — typed -> str —
and the next line picks the harness from it, which would raise on an empty one.
AgentCall.start(model=model) is fifty lines further on, past provisioning. So
the row is only ever written once the model is known, and Agent.model is a
required field with no default.

• model: Mapped[str], and AgentCall.start(model: str)
• migration c71b3d95e802, no backfill for the same reason agent had none:
  a guessed model would land in the cost breakdown
• seed_call and seed_note_agent_run take model: str

Two tests pinned the case that can't happen

test_usage_today_counts_calls_whose_model_isnt_a_current_harness seeded both
"claude-opus-4-5" and None. Its own comment states the real requirement —
model ids churn on deploys, so a call finished earlier today can carry an id no
harness claims any more — and that half stays, along with the unattributed
bucket it lands in. The None half went.

test_agent_call_artifact_layout seeded model=None to show "the layout holds
with the model unresolved". The layout's actual rule is that it is the same
whichever harness ran the call, which is what it now says.

The usage panel keeps its unattributed bucket either way: a pinned or churned
id that no picker list claims is a real case, and money spent must not vanish
from the display.