fix(ai): correct off-by-one in array form of mock model helpers

[serhiizghama]

Problem

MockLanguageModelV3 / MockLanguageModelV4 and MockEmbeddingModelV3 / MockEmbeddingModelV4 accept an array of results to script multi-call test scenarios (e.g. tool-call → tool-result → final text for ToolLoopAgent). The array form silently misbehaved: the first call returned the second entry, and the last call returned undefined.

The cause is the indexer running after the call has already been pushed onto the call list:

this.doGenerate = async (options) => {
  this.doGenerateCalls.push(options);          // length is incremented…
  if (Array.isArray(doGenerate)) {
    return doGenerate[this.doGenerateCalls.length]; // …then used as the index
  }
  ...
};

The same off-by-one is in doStream for both V3 and V4 and in doEmbed for both embedding mocks.

Fixes #15141.

Solution

Index with length - 1 after the push so the n-th call returns the n-th array entry, matching the obvious user expectation.

Testing

Added mock-language-model-v3.test.ts, mock-language-model-v4.test.ts, mock-embedding-model-v3.test.ts, mock-embedding-model-v4.test.ts covering both doGenerate / doStream / doEmbed array forms. Verified the new tests fail against main and pass with the fix. Existing wrap-language-model and wrap-embedding-model tests continue to pass (they use empty arrays, so behavior is unchanged for that path).

[lgrammel]

Fixed via #15154