feat: report what a run spent on the model

[mohammadp1001]

The problem

Nothing in Alphoryn reported what a run cost. Establishing the 2026-08-13 figure
meant querying Cloud Logging by hand, and the answer was uncomfortable:

Model spend	~$2.90
Trading P&L that day	-$0.59

The run cost more than the entire P&L swing, in both directions. The manual dig
also turned up the thing that actually matters: 94% of output tokens were
reasoning tokens, and reasoning bills at the output rate - 8x the input rate.
Thinking is the cost.

The fix

Both agents accumulate per-call usage and emit TOKEN_USAGE telemetry. The run
prints a per-agent and per-run summary when it ends:

  gemini-2.5-pro: Model usage: 116 calls, 979,824 in (576,558 cached) / 207,375 out, 99% reasoning  ->  ~$2.76
  gemini-2.5-pro: Model usage: 14 calls, 8,581 in (0 cached) / 12,212 out, 0% reasoning  ->  ~$0.13
Run Model usage: 130 calls, 988,405 in (576,558 cached) / 219,587 out, 93% reasoning  ->  ~$2.89

That $2.89 is the same number the manual Cloud Logging query produced for
run-4, which is how this was checked.

Three decisions worth flagging

Output is total - prompt, not candidates_token_count. Whether that field
includes thought tokens has changed between API versions; the subtraction cannot
drift, because everything that is not prompt bills at the output rate.

Usage is recorded before the failure paths. A session that failed still spent
what it spent - the live run burned three retries on a session that produced
nothing, and that is precisely the spend you want visible. The run summary prints
from the loop's finally for the same reason: both live runs so far ended by
being killed, and an aborted run is when you most want the number.

format_summary takes a cost, not a model, so a total spanning two
differently priced models is summed per-model by the caller. Pricing a mixed
total at one model's rate is how a cost report starts lying. An unpriced model
reports its tokens and says cost unknown rather than guessing.

usage_from_event is a single shared reader for both agents, deliberately - the
last time this kind of event-reading logic was written twice it drifted and cost
the feedback agent a month of silent failure (#178).

What this does not do

It does not reduce the cost. Note for whoever picks that up: do not reach for
disabling include_thoughts - per alphoryn/agents/thinking.py, 2.5 models
reason either way, so you would still pay for the tokens and merely lose the
trace. The real levers are a thinking budget cap or a cheaper model for
main_agent.

Verification

• ruff check alphoryn/ tests/ clean
• 734 tests, 100% coverage (usage/tokens.py 39 stmts / 6 branches, all covered)
• Reproduced the CI environment: GOOGLE_APPLICATION_CREDENTIALS=/nonexistent/adc.json python -m pytest -q