fix(sdk): take max of OpenRouter cost and upstream_inference_cost

[nordicnode]

Summary

• user-facing credits were computed as usage.cost + cost_details.upstream_inference_cost at all three cost sites (sdk/src/impl/llm.ts streaming reportCost, generateText, and structured paths)
• usage.cost is the total amount charged and already includes the upstream portion reported separately in cost_details, per OpenRouter's usage accounting docs — summing them roughly doubles credits on normal (non-BYOK) OpenRouter routes, which lines up with the "extreme inflation" reports on Discord
• take the max instead, matching the server-side ledger's documented semantics (the Solar Pro 4 BYOK note in common/src/constants/freebuff-models.ts: cost = 0 there and upstream carries the real spend, so max is correct for both shapes)
• add a regression test asserting exact credits for cost=0.01, upstream=0.02 (fails under the sum: 3 vs 2)

Fixes #1164

Validation

Re-run on the new main after the history rewrite:

• bun test sdk/src/impl/__tests__/usage-receipts.test.ts — 5 pass, 0 fail
• bunx tsc --noEmit -p sdk/tsconfig.json clean

---

Replaces #1165 — auto-closed when the repo history was force-pushed (the original commits no longer exist in the new history). Rebased onto the new main per the maintainer's request; picking the review up here.

[codebuff-team]

Good instinct and scoped correctly — sdk/src/impl/llm.ts is fair game, and you touched all three cost sites (streaming reportCost, generateText, structured) consistently rather than fixing only one. The added regression test in usage-receipts.test.ts is a real behavioral test (drives an SSE stream through promptAiSdkStream and asserts the emitted credit amount), not just a unit check on a helper, which is the right level for this kind of bug.

One thing worth flagging for the maintainer to weigh: Math.max(cost, upstream) is a heuristic that happens to work for the two cases you cite (normal OpenRouter routes where cost already includes everything and upstream_inference_cost is 0/absent, and BYOK routes where cost is 0 and upstream_inference_cost carries the real spend) but it silently drops any genuine OpenRouter markup fee in a BYOK request where both cost and upstream_inference_cost are simultaneously nonzero — i.e. if OpenRouter ever charges a small fee on top of the upstream cost for BYOK, max will undercount vs the true sum. That's a much smaller and safer failure mode than the current doubling bug, but it's worth the maintainer confirming against OpenRouter's actual accounting semantics for BYOK-with-fee before this ships, rather than assuming max is universally correct.

Overall: focused, in-scope, has a test, addresses a real reported issue (#1164) with a concrete before/after fixture. Worth porting, with the max-vs-sum edge case called out during review.