Tune parallel recovery scale-up increment to 0.15

[hemanth-asirvatham]

Motivation

• Clarify the first observed connection error message to help users diagnose possible network instability or bandwidth limitations.
• Make parallel-worker recovery scale-ups less frequent but larger by increasing the increment multiplier to 0.15 so recoveries are more meaningful.

Description

• Updated the first-connection error message in src/gabriel/utils/openai_utils.py to mention that it "may be due to network instability or bandwidth limitations."
• Tuned the parallel recovery logic in src/gabriel/utils/openai_utils.py by keeping the success_threshold at ceil(concurrency_cap * 2.0) and changing the scale-up increment calculation to max(1, int(math.ceil(max(concurrency_cap * 0.15, 2)))).

Testing

• No automated tests were run for these changes.

---

Codex Task

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}