fix: [#836] address re-review — rename misleading test + document negative-cost follow-up

Reviewer caught that the test 'tokens.input is never negative even with
inconsistent provider counts' is the OPPOSITE of what it actually
verifies. The test pins that the algebraic invariant
input + cache.read + cache.write === inputTotal holds when input IS
negative — not that input avoids being negative.

Renamed to 'invariant holds even when tokens.input goes negative on
inconsistent provider counts' which matches the test body. Expanded
the comment to note the downstream cost-calculation implication
(negative input × positive cost rate produces a negative cost
contribution) as a pre-existing follow-up concern.

No code behavior change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: anandgupta42

packages/opencode/test/session/session-getusage.test.ts

@@ -164,13 +164,23 @@ describe("Session.getUsage — provider edge cases", () => {
     expect(result.tokens.inputTotal).toBe(800 + 2000 + 600)
   })
 
-  test("tokens.input is never negative even with inconsistent provider counts", () => {
+  test("invariant holds even when tokens.input goes negative on inconsistent provider counts", () => {
+    // Renamed from "tokens.input is never negative" — the previous name was
+    // the OPPOSITE of what this test verifies. The test pins that the
+    // algebraic invariant `input + cache.read + cache.write === inputTotal`
+    // holds even when input is negative.
+    //
     // Hypothetical: OpenAI returns inputTokens=1000 but cachedInputTokens=2000
     // (inconsistent — should never happen but providers occasionally surface
-    // weird numbers). Verify the subtraction doesn't underflow into negative
-    // territory; safe() clamps via Number.isFinite but does NOT clamp
-    // negatives. Document the current behavior so a future refactor that
-    // changes it is forced through this test.
+    // weird numbers). Verify the invariant holds even though `safe()`
+    // clamps non-finite values but does NOT clamp negatives. Documents the
+    // current behavior so a future refactor that changes it is forced
+    // through this test.
+    //
+    // Note: negative tokens.input flows into cost calculation as a
+    // negative contribution. Pre-existing concern — not introduced by
+    // this PR. Tracked as a follow-up to either clamp at zero or accept
+    // the offset.
     const result = Session.getUsage({
       model: fakeModel("@ai-sdk/openai"),
       usage: {