fix: anchor retry backoff reset to the first event of each connection

[tanderson-ld]

Summary

retryResetIntervalMillis is documented as resetting the backoff once the stream "has remained active for some amount of time," but the implementation re-anchored the internal "good since" timestamp on every received event. That measured idle-time-before-failure rather than connection health: a chatty stream that had been active for hours never qualified for a reset (its last event was always moments before the failure), while a stream that went silent for the interval and then failed did. Tracked as SDK-2846.

Fix

The "good since" timestamp is now anchored once per connection, at the first event received on that connection. A connection that has been delivering data for at least retryResetIntervalMillis resets the backoff to the initial delay on its next failure. A connection that fails before delivering any event — including one that only sent comment heartbeats — does not count as healthy and leaves the backoff progression intact.

The README's two descriptions of this option contradicted each other (line 93 described connection-anchored semantics, the example comment described last-event semantics); both now describe the fixed behavior.

Behavior change

This changes released retry-reset timing for all consumers: long-lived active connections now correctly reset the backoff (previously they never did), and briefly-idle-then-failed connections no longer do. No API changes.

Testing

• New test: a connection that delivers events continuously past the reset interval and then drops must reset the delay to the initial value. Verified to fail against the previous implementation.
• New test: a connection that stays open past the reset interval sending only comment heartbeats must not reset the backoff.
• Full suite: 90 passing, standard clean. New tests stable across repeated runs.

---

Note

Overview
Fixes retryResetIntervalMillis so backoff reset matches the documented “connection has been active” semantics instead of measuring idle time since the last event.

receivedEvent now calls setGoodSince only once per successful stream (first parsed event), with goodSinceAnchored cleared when a new HTTP 200 connection starts. Long-lived streams that keep delivering events can reset exponential backoff after the interval; connections that only send comment heartbeats or fail before any event still do not count as healthy.

README wording for this option is aligned with the implementation. Tests add verifyDelaysWithHandler plus scenarios for eventful long connections vs comment-only heartbeats.

^{Reviewed by Cursor Bugbot for commit bad4092. Bugbot is set up for automated code reviews on this repo. Configure here.}