fix(gateway): reconcile enforced receipts from failed executions

[drewstone]

A sandbox error could discard an authoritative receipt emitted after paid inference. Payment cleanup then retained the quote despite having measured usage.

Validate and forward enforced usage before terminal protocol errors, then settle the measured amount while preserving OpenAI errors and failed A2A tasks. Missing or unenforced receipts retain ownership under the existing recovery policy. Settlement acknowledgement failures preserve the durable measured receipt. Ordinary finalization failures keep their existing recovery behavior.

Validation: 498 tests pass on Node 24.18.0 and 22.22.2; typecheck and build pass. New coverage exercises OpenAI JSON/SSE, both A2A methods, real SQLite capped-key reservations, acknowledgement-loss recovery, and published agent-app 0.47.6 with the maintained Runtime producer, persisted chat route, and gateway bridge. Removing the failed-receipt handoff fails the composed regression.

All remote inference and payment effects use synthetic callbacks. The tests establish local accounting and protocol behavior, not external payment transfer or live provider enforcement. Version 0.11.1 contains this correction.

[tangletools]

Independently reviewed the failed-receipt dispatch, settlement, and A2A recovery changes at this exact head. Validated usage is forwarded before the terminal execution error; durable settlement preserves measured receipts across acknowledgement loss. A2A clears only reconciled recovery markers and retains failed task status. Reviewed the composed test against registry agent-app 0.47.6 and the receipt-removal mutation evidence. Both configured CI runtimes passed 498 tests. These are synthetic transport/payment fixtures, not a claim of live provider or paid-customer proof. No blocking findings.