Meter what a run and a gate actually cost, so the window ceilings hold

[ranjitha13g]

daily_budget and daily_triage_budget could never fire: both accumulated 0.00 forever, and the morning report's two cost lines were permanently zero. The governor was correct; the engine fed it zero, because it read money from fields only its own test doubles wrote.

Three facets, one root cause. engine.py read result['spend_usd'], but AgentRuntime.run() reports budget.spent and never emitted that key. _spend_of read reply['metered_calls'], but GatewayClient.complete -- the callable the HTTP event route passes in -- carries no meter at all. And it looked for cost_usd/usd, while BudgetedGateway writes Charge.as_dict(), whose money field is cost.

Fix: runtime.run() publishes spend_usd at the top level; the engine falls back to budget.spent; _spend_of also reads cost, and prices an unmetered reply's own tokens against config/pricing.yaml via the existing Pricing.cost(). No new price appears in Python.

The runtime double now builds its return from a real RunBudget.snapshot() instead of a hand-written dict, so this class of drift fails the suite rather than passing it. Reverting the two source files fails 5 tests, two of which already shipped green.

357 passed, up from 353.

[theschoolofai]

Session 16 — graded

Score: 0.

Duplicate of two earlier findings at once: metering the gate is #7 and metering the run is #10, both filed 2026-08-10 against this one's 08-13. Combining them into one PR does not create a new finding.

[theschoolofai]

Session 16 — regraded ✅

Score: +100. The verdict below stands: this is a duplicate of an earlier filing (or, for glc_v5 #23, an enhancement rather than a defect), and it is recorded as such.

What has changed is the credit. On review, the work here was genuinely done: the bug was found independently, the reproduction is real and the fix is sound. Losing a race you had no way of seeing is not a reason to earn nothing, so this is credited at the full 100 even though it does not carry the first-to-file claim.

---

Original assessment, unchanged:

Duplicate of two earlier findings at once: metering the gate is #7 and metering the run is #10, both filed 2026-08-10 against this one's 08-13. Combining them into one PR does not create a new finding.