[1828703450qqcom-ux]

fix(adk): bypass retry for mid-stream graph interruptions

Fixes #1446

Summary

• The Generate retry path already checks compose.ExtractInterruptInfo(err)
  and returns immediately for interrupt errors. The stream retry paths
  (both ShouldRetry and legacy) did not — a graph interruption delivered by
  a model stream was treated as a retryable error, and the next attempt's
  output replaced the original interruption.
• Fix: add the same interrupt check to the stream error handling in both
  streamWithShouldRetry and streamLegacy.

Reproduction

From the issue, with a model whose first Stream() call emits one chunk then
an interrupt error, and a second call that succeeds:

• On main: the retry layer calls the model a second time and returns the
  second attempt's output. The interruption is lost.
• After the fix: the original interruption error is preserved.

Fix

adk/retry_chatmodel.go — two hunks, both adding the same pattern after
stream error consumption and before the retry decision:

if streamErr != nil {
    if _, ok := compose.ExtractInterruptInfo(streamErr); ok {
        // bypass retry, preserve the original interruption
        ...
        return nil, streamErr
    }
}

1. streamWithShouldRetry: added after typedConsumeStream and before
   ShouldRetry is called.
2. streamLegacy: added after consumeStreamForError and before
   isRetryAble is called.

The Generate path (both ShouldRetry and legacy) already has the same check
and is unchanged.

Validation

• Full adk suite: go test ./adk/ -count=1 -timeout 120s → ok (49s).
• All retry/stream tests pass (go test ./adk/ -run "Retry|Stream" -count=1).
• go vet clean.

Scope

• 1 file: adk/retry_chatmodel.go (two hunks in stream retry paths).
• No change to the Generate retry path, retry policy, backoff computation,
  or any other logic beyond interrupt bypass.