Discard stale throttled pending content when NewEntry re-sends fresh entries

[lukemarsden]

Problem

The text block immediately before a long-running tool call renders truncated in Helix (e.g. "The board" instead of the full paragraph) for the entire duration of the tool call, while Zed shows the full sentence. Fourth recurrence of this bug (prior attempts: spec tasks 001655, 001694, 002552, 003117).

Root cause

A regression introduced by 73f9af2162 (the previous fix attempt): the NewEntry handler re-reads and sends every current-turn entry with complete content, but no longer clears the streaming throttle's pending_content for the preceding text entry. The stale mid-sentence snapshot is then sent after the fresh full content — by the trailing flush timer or by the next entry's flush-others step — and Helix's last-write-wins per message_id keeps the truncated text until the end-of-turn flush.

Fix

• discard_pending_for_thread() drops pending throttle state (and defuses the armed trailing flush timer) right before the NewEntry handler's fresh re-send.
• Deletes the dead flush_stale_pending_for_thread left behind by 73f9af2162.
• Regression test pinning the ordering, plus a control test proving the clobber occurs without the discard.
• go mod tidy for the e2e test-server module (needed against current helix deps).

Testing

• cargo test -p external_websocket_sync: 81 passed, 0 failed.
• Live in a helix-in-helix stack built from this branch: 14+ text→tool_call boundaries across 3 spec tasks — every pre-tool-call paragraph rendered complete mid-turn, with a 1s DOM sampler showing zero regressions over a 5-minute window.
• NOT run: the dockerized e2e LLM rounds (no Anthropic credential available in the dev sandbox).

Release Notes:

• N/A

🤖 Generated with Claude Code

---

🔗 Open in Helix

📋 Spec:

• Requirements
• Design
• Tasks

🚀 Built with Helix