Requests kept failing but AIC consumption kept increasing

### Describe the bug

I had a terrible experience today where requests kept failing with `Response was interrupted due to a server error. Retrying...` and `Request failed due to a transient API error. Retrying...` all the while my AIC consumption kept increasing

I was using GPT 5.4 with a 400k context. Everything worked perfect in the first hour during a `/grill-me` session, then I switched to autopilot for the implementation. It worked for a while then it hit a few hiccups so I left the AI on its own while I take a lunch break. Came back to a session that consumed half of the allocated credits for a month, and almost nothing to show for it

In the screenshot above, the session has taken up 168k / 400k of context. I saw it starting to fail with transient errors so I went for lunch and left the AI to work.

An hour later I get this:

### Affected version

1.0.62

### Steps to reproduce the behavior

Do an intensive planning session using the [`/grill-me`](https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md) skill by [Matt Pocock](https://github.com/mattpocock/skills)

Switch to `autopilot` for the implementation

Watch it burn a little then go for lunch

### Expected behavior

I would expect either that the tool gives up and stops trying, or at least that the AIC consumption doesn't increase.

I shouldn't have to pay to use a faulty system

### Additional context

This was on a Github Enterprise account with EU data residency

Reproduziere dies in Version 1.0.62 mit einer intensiven /grill-me-Sitzung, gefolgt von autopilot, und beobachte dann das Verhalten, nachdem Wiederholungsversuche bei transienten Fehlern oder Serverfehlern beginnen. Verfolge die Pfade für die Wiederholungsbehandlung und die AIC-Abrechnung; abgeschlossen ist die Aufgabe, wenn das Tool die Wiederholungsversuche beendet oder anderweitig verhindert, dass bei wiederholt fehlschlagenden Anfragen weiterhin AIC verbraucht wird.

Vom Indexierungsmodell aus dem Issue-Text verfasst.