fix(gateway): sanitize lone surrogates before platform delivery

[AlexFucuson9]

Fixes #55309

_sanitize_gateway_final_response() redacted secrets and classified provider errors but did not strip lone surrogate code points (U+D800-U+DFFF). These are invalid in UTF-8 and crash the utf-16-le length check in Telegram delivery, aborting the send for the entire reply.

The codebase already had _sanitize_surrogates() in agent/message_sanitization.py (used for history storage), but it was never applied to the delivered copy.

Fix: Apply _sanitize_surrogates() as the final step in _sanitize_gateway_final_response() so all chat surfaces receive surrogate-safe text.

Reproduction (before fix):

content = "Here is your answer \ud800 done"
utf16_len(content)  # UnicodeEncodeError

After fix: lone surrogates are replaced with U+FFFD before reaching platform adapters.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #55310 (earliest open canonical fix for #55309, paired with the issue by its author). Same _sanitize_surrogates() application in _sanitize_gateway_final_response() (gateway/run.py). Related: #55309 (issue), #55143/#55144 (Signal sibling of the same lone-surrogate class).

[tonydwb]

LGTM. Strips lone surrogate code points that crash utf-16 encoding in platform length checks. Clean fix.

[teknium1]

Thanks for tracing the non-streaming final-response boundary; current main still returns ordinary chat replies unsanitized at gateway/run.py:434-437, so that portion addresses a real defect.

Problems

• Streaming bypasses this helper: gateway/run.py:18123-18125 forwards raw deltas to GatewayStreamConsumer.on_delta, which queues them unchanged at gateway/stream_consumer.py:370-379 and later edits through the adapter at gateway/stream_consumer.py:1771-1775. Telegram measures edit content with strict UTF-16 encoding at plugins/platforms/telegram/adapter.py:3980, so a lone surrogate can still abort delivery before this final-response path runs.
• The diff changes only gateway/run.py; it adds no regression test. tests/gateway/test_telegram_noise_filter.py:79-243 has final-response coverage but no lone-surrogate case.

Suggested changes

• Retain the final-response sanitization and sanitize deltas at the stream-consumer input boundary; add non-streaming and streaming Telegram regressions that prove the downstream UTF-16 measurement accepts the delivered content.
• Keep raw programmatic surfaces unchanged, consistent with gateway/run.py:421-427.

Automated hermes-sweeper review.

[teknium1]

Please add a regression test for this final-response boundary, and cover the streaming path separately: raw deltas bypass this helper via gateway/run.py:18123-18125 → GatewayStreamConsumer.on_delta, then reach adapter edits before the final sanitizer can run.