fix: handle a reply cut off part-way through an A2UI message

[brianegan]

Summary

When Gemini's recitation filter stopped a reply, the example chat showed half-written A2UI JSON as prose, then an opaque "Internal server error". The recitation stop sends a chunk with empty content, and genkit_google_genai crashes on it with a null check that Genkit reports as INTERNAL. The plugin bug belongs upstream. This PR makes the package and the example handle a reply that is cut off.

Changes

• A2uiParserTransformer reports an A2UI message that is still unfinished when the stream ends as an A2uiValidationException, so it arrives as a GenUiError. This applies whether the stream ended cleanly or failed. The leftover counts only when it opens with a ```json fence, or with a brace or fence that names a message key. A stray brace or an unrelated code block still shows as prose.
• The example shows "The model stopped before finishing. Try again, or ask a different way." for a failed turn and logs the raw failure to the browser console. Whatever the reply managed to produce stays in the transcript.
• The example's chatHandler logs a failed turn to stderr with its cause and stack trace. A model failure doesn't throw out of a Genkit agent. It ends the turn with AgentOutput.error, so the wrapper reads the error there, before genkit_shelf hides it behind a 500.
• The example prompt asks the model to write in its own words and not quote reference text, which makes recitation stops less likely.
• The README lists a message cut off when the reply ended as one kind of GenUiError.

Testing

The best check is the real failure. Run the example with GEMINI_API_KEY set, and send a prompt that asks for stock text word for word, for example: "Quote verbatim the Google knowledge panel description of Paris that begins: Paris, France's capital, is a major European city and a global center for art, fashion, gastronomy and culture. Continue it word for word." Recitation doesn't trigger on every run, so it can take a few tries.

When it does trigger, the model's opening sentence stays in the bubble with no JSON after it, a red line says the reply ended part-way through an A2UI message, and the plain failure message appears under the transcript. The server terminal should show Chat turn failed: INTERNAL: Google AI Error: Null check operator used on a null value with a stack trace that points at fromGeminiCandidate in genkit_google_genai.

Without a key, test/transport/parser_test.dart covers the truncation cases, example/test/browser/chat_browser_test.dart covers the failed-turn UI, and example/test/round_trip_test.dart covers the server log.

Gates

Gated 496db48 against origin/main@5350ed7.

• review: skill code-review on fable, passed, agent-reported
• test: command, passed in 84.2s
• example-test: command, passed in 16.1s
• browser-test: command, passed in 6.6s

Reviewed against the plan recorded on the issue.

Closes #25