feat: conform FDv1 streaming and polling data sources to the RETRY spec

[tanderson-ld]

Summary

Wires the reusable retry controller (added to @launchdarkly/js-sdk-common in #2045) into the Node server SDK's FDv1 streaming and polling data sources, bringing them into conformance with the RETRY spec. Streaming injects the controller through the js-eventsource retryDelayStrategy seam; polling drives the controller from its own scheduled loop. This replaces the old recoverable / permanent-stop model.

Scope is FDv1 streaming + polling in the Node server SDK only. FDv2 (the composite data source and its synchronizers) is intentionally untouched and will be conformed in a future effort.

Draft / dependency note. The streaming seam ships in an unreleased launchdarkly-eventsource; the committed pin here stays at 2.2.0, so CI does not yet exercise the injected controller end-to-end (2.2.0 ignores the option and falls back to its built-in backoff). The pin bump to the release carrying the seam, and the long-running contract-test verification, are follow-ups — this PR is up for review of the wiring and behavior.

Behavior changes (operator-facing)

• No data-source failure is terminal. Previously-fatal HTTP statuses (401, 403, and other 4xx) no longer permanently stop the data source; they retry indefinitely at the extended-regime cadence (5 min → 1 hr). Normal/transient failures retry at the fast cadence (1 s → 30 s, streaming) as before.
• A bad SDK key no longer fast-fails initialization. waitForInitialization() resolves on success and rejects only on its own timeout or on close(); it no longer rejects immediately with "Authentication failed…" on a 401. A misconfigured client keeps retrying and stays uninitialized until the timeout.
• Malformed/empty data is recovered from, not fatal. Malformed stream JSON / empty payloads record a normal failure and re-establish the stream (self-initiated close + recreate after the backoff); malformed poll JSON records a normal failure and polls again on schedule.
• Data-source failures surface as logs only. Consistent with how the SDK has always treated recoverable/interrupted failures, and with the Go/Java/.NET fleet: a fixed "will retry" message (warn for normal, error-level for unexpected/invalid-data), with the reconnect delay in the streaming onretrying info line. The 'error'/'failed' events and init-rejection remain reserved for genuinely terminal conditions (today, only the untouched FDv2 composite reaches them).
• Polling cadence is completion-anchored (effective period ≈ interval + poll duration), matching the fleet.

Versioning: feat: / minor for all affected packages — matching how Go (7.16.0), Java (7.16.0), and .NET (8.17.0) each shipped RETRY conformance (minor, behavior changes documented in the PR, not marked breaking).

What's here

• packages/shared/common — EventSourceRetryDelayStrategy type + optional retryDelayStrategy on the platform EventSourceInitDict (the injection seam; initialRetryDelayMillis/retryResetIntervalMillis made optional since they're unused under injection).
• packages/shared/sdk-server — StreamingProcessor and PollingProcessor rewired to the controller (classification, always-retry, malformed-data recovery via a guarded close-and-recreate, completion-anchored polling). LDClientImpl is intentionally left unchanged; the reserved 'error'/'failed'/init-reject path is untouched.
• packages/sdk/server-node — NodeRequests passes the strategy through; contract-test capabilities retry-conformance-fdv1-streaming / -polling; a nightly workflow (server-node-nightly.yaml) that runs the long-running RETRY suite (schedule + manual dispatch, no push trigger).

Testing

Unit suites green: @launchdarkly/js-server-sdk-common (1053 pass) and @launchdarkly/node-server-sdk (70 pass), covering the classification ladders (via the controller), the streaming close+recreate on malformed data (including a guard against a duplicate reconnect and a deserialization that throws), polling recovery, stop()-cancels-a-pending-retry interruptibility, status-less transport-error classification, single-failure accounting for a self-initiated restart, the reserved terminal 'error'/'failed'/init-reject path, and the restored onretrying delay log. The behavioral RETRY conformance itself is validated by the harness in the nightly (long-running), which is a post-merge / manual-dispatch activity.

Part of SDK-2790.

---

Note

Overview
Brings FDv1 streaming and polling in the Node server SDK in line with the RETRY spec by driving reconnect/poll timing from shared RetryState instead of permanent vs recoverable stops.

Streaming injects that state through a new EventSourceRetryDelayStrategy on the platform EventSourceInitDict (optional initialRetryDelayMillis / retryResetIntervalMillis; browser defaults when omitted). Polling schedules the next request from the same controller with completion-anchored delays. 401/403, transport errors, and bad JSON are logged and retried (extended cadence for “unexpected” HTTP); they no longer surface data-source 'error' events or stop the loop. Malformed stream payloads trigger a guarded close-and-reconnect; poll deserialization failures keep polling.

Also bumps launchdarkly-eventsource to 2.3.0, preapproves it in .yarnrc.yml for the npm age gate, advertises retry-conformance-fdv1-* contract capabilities, and adds a nightly GitHub workflow for the long-running RETRY harness (--enable-long-running-tests).

^{Reviewed by Cursor Bugbot for commit 54ba5dd. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

@launchdarkly/js-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 29338 bytes
Compressed size limit: 29500
Uncompressed size: 141229 bytes

[github-actions]

@launchdarkly/js-client-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 25437 bytes
Compressed size limit: 44000
Uncompressed size: 165420 bytes

[github-actions]

@launchdarkly/js-client-sdk size report
This is the brotli compressed size of the ESM build.
Compressed size: 32669 bytes
Compressed size limit: 34000
Uncompressed size: 117050 bytes

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9fbd27a. Configure here.}

[joker23]

We should bump the eventsource package version.