Fix Infinity overflow in retry backoff calculation

[pavankumar-vh]

Overview

Fix Infinity overflow in retry backoff calculation in common/src/util/promise.ts.

Bug Description

The function used Math.pow(2, attempt) which can overflow to Infinity if attempt is large. Then Math.round(Infinity * jitter) returns Infinity, and setTimeout with Infinity would never fire, causing the retry to hang forever.

Fix

Added Number.isFinite() check to cap the delay at Number.MAX_SAFE_INTEGER.

Testing

No existing tests for this function, but the fix prevents infinite hangs.

Files Changed

• common/src/util/promise.ts - Added overflow protection

Scope

This change only touches common/ which is an approved contribution area per the Contributing Guide.

[codebuff-team]

Thanks for tracking this down — the overflow scenario is real: with enough retry attempts, Math.pow(2, attempt) does hit Infinity, and Math.round(Infinity * jitter) is NaN (not Infinity as stated, since Infinity * jitter where jitter is finite is Infinity, but the multiplication itself is fine — actually Infinity * 0.9 is Infinity, so Math.round gives Infinity, confirmed). setTimeout with Infinity per the HTML/Node spec clamps to a 1ms delay or fires immediately in most engines, not "hangs forever" — worth double-checking that claim, as it changes the severity of the bug you're describing.

On the fix itself: capping at Number.MAX_SAFE_INTEGER (~9e15 ms) is not really a usable delay either — setTimeout will clamp it to its internal max (2^31-1 ms, ~24.8 days) and the retry will still take an absurdly long time to fire. A more useful fix would cap baseDelayMs at some sane maximum backoff (e.g., a maxDelayMs parameter or a fixed ceiling like 30s), not just at whatever avoids Infinity/NaN.

This also has no accompanying test, and the repo doesn't have existing tests for withRetry per your own note — a maintainer would likely want at least a unit test asserting the delay is capped for large attempt counts before porting this.

Right instinct, but the numeric ceiling chosen doesn't actually solve the practical problem of overly long retry delays, and it needs a test to be portable as-is.