Expose provider retry progress and attempt counts to ACP parents

[danielkov]

Summary

Kit now retries transient openai-subscription failures internally, but the parent sees no bounded progress while those retries are occurring. A long-running subagent(...) call can therefore appear stalled until it succeeds or exhausts its retry budget.

Split from #35 after the overload termination itself was fixed by #32 and released in v0.1.98.

Current behavior

The provider calculates the retry attempt and wait, then sleeps without emitting a runtime progress event. This is especially opaque when subagent(...) runs inside compose(background=true).

Desired behavior

Emit sanitized, bounded retry progress that a parent and UI can observe. Each update should include at least:

• provider route;
• current attempt number;
• reason category without raw provider payloads;
• next delay; and
• cumulative retry/backoff time.

Updates should be rate-limited or coalesced so sustained provider failures do not flood transcripts. The terminal result should remain the single success or failure result.

Suggested coverage

• A fake provider fails N times and succeeds: assert observable attempt updates and one final success.
• An always-failing provider: assert bounded updates followed by exhaustion.
• Exercise foreground subagents and compose(background=true).
• Assert updates contain no provider message body, credentials, prompts, or customer identifiers.

Environment

• Originally observed with Kit 0.1.96 on Darwin 25.6.0 arm64.
• Confirmed as a remaining gap on Kit 0.1.105 on Darwin 25.5.0 arm64.
• Surface: ACP; provider route: openai-subscription.

Related work

• ACP subagents terminate on repeated openai-subscription server overloads #35
• fix(provider): overhaul openai-subscription retry policy #32
• Add a centralized activity view for background compose calls and subagents #36 can present these updates in its centralized activity view, but this issue owns producing the retry event.