fix(gateway): use code-point budget for fallback chunk split boundary

[sprmn24]

What does this PR do?

_split_text_chunks() correctly calculates _cp_budget (code-point boundary) via _custom_unit_to_cp() for adapters using custom length functions (e.g. UTF-16 unit counters). However, the fallback path when no suitable newline is found was using limit (raw platform units) instead of _cp_budget (code points) as the split boundary.

For adapters measuring UTF-16 units, emoji and other astral-plane characters count as 2 units each, so slicing at limit instead of _cp_budget can produce chunks that substantially exceed MAX_MESSAGE_LENGTH and fail delivery.

This PR fixes both the split boundary and the too-short newline guard to consistently use _cp_budget.

Related Issue

None

Type of Change

• 🐛 Bug fix

Changes Made

• gateway/stream_consumer.py L977-978: changed split_at < limit // 2 → split_at < _cp_budget // 2 and split_at = limit → split_at = _cp_budget in _split_text_chunks()

How to Test

1. Configure a UTF-16 length adapter (e.g. Telegram)
2. Send a message containing emoji-heavy text that exceeds the platform limit
3. Confirm all returned chunks satisfy len_fn(chunk) <= limit

Checklist

• I have read the Contributing Guide
• Commit messages follow Conventional Commits format
• I have checked for duplicate PRs
• This PR is scoped to a single fix
• pytest tests/ -q passed
• Platform: Windows 11 + WSL2 Ubuntu

Documentation & Housekeeping

• README/docs update: N/A
• cli-config.yaml.example: N/A
• CONTRIBUTING.md/AGENTS.md: N/A
• Cross-platform impact: N/A
• Tool description/schema: N/A

Screenshots / Logs

N/A

[alt-glitch]

This was generated by AI during triage.

Duplicate of #42523 — same _cp_budget fix in the same _split_text_chunks() no-newline fallback (UTF-16 unit vs codepoint slice boundary). #42523 is the earliest open PR for this fix and is the canonical one. Also related to #55843 (same fix + regression tests, carries issue #55844). Part of the UTF-16 split cluster (#11467/#23455/#11170/#11522).

[teknium1]

Thanks for isolating the fallback boundary mismatch. The proposed _cp_budget slice is correct for _split_text_chunks(): _custom_unit_to_cp() explicitly returns the largest code-point offset fitting the custom length budget (gateway/platforms/base.py:167-183).

Problems

• The same unsafe fallback remains in the existing-message streaming overflow loop. It computes _cp_budget at gateway/stream_consumer.py:701-704, but falls back to _safe_limit as a Python slice offset at gateway/stream_consumer.py:705-707.
• This PR adds no UTF-16 regression coverage. The existing fallback test (tests/gateway/test_stream_consumer.py:778-805) uses a MagicMock/default len() path and cannot assert the custom-unit chunk invariant.

Suggested changes

• Add a direct UTF-16 no-newline regression asserting every _split_text_chunks() result satisfies utf16_len(chunk) <= limit and round-trips the input.
• Apply the same _cp_budget boundary correction to the sibling streaming-edit fallback at gateway/stream_consumer.py:705-706.

Automated hermes-sweeper review.

[teknium1]

Please add a direct UTF-16 regression for this branch: an emoji-only, no-newline input should produce only chunks whose utf16_len is within limit. The current fallback tests exercise default len() behavior and would not catch this unit/code-point mismatch.