feat: allow a custom retry delay strategy to be provided

[tanderson-ld]

Summary

Adds a retryDelayStrategy init option that lets the application take full control of reconnection timing, and hardens handling of the SSE retry: field.

• retryDelayStrategy: an object implementing the same three-method contract as the built-in strategy (nextRetryDelay(currentTimeMillis), setGoodSince(goodSinceTimeMillis), setBaseDelay(delayMillis)). When provided it fully replaces the built-in delay behavior, so initialRetryDelayMillis, maxBackoffMillis, jitterRatio, and retryResetIntervalMillis have no effect (documented in the README). Behavior is unchanged when the option is absent.
• retry: field hardening: the value must now be entirely ASCII digits, per the SSE specification — previously parseInt leniency meant retry: 12abc was applied as 12, and a negative value produced an immediate-reconnect tight loop — and accepted values are capped at one hour, so an enormous reconnection time cannot behave like a permanent stop. Existing semantics are otherwise preserved (the value replaces the base delay, restarts the backoff progression, persists until the server sends another, and never changes the maximum).
• EventSource.supportedOptions includes the new option.

This is the library half of the Node server SDK's RETRY-conformance work (SDK-2790): retry policy moves into the SDK behind this seam, while the library keeps the connection and timer mechanics it already has (including cancellable reconnect timers and one-failure-per-connection idempotence).

Testing

• Nine new tests: an injected strategy drives reconnect delays; the built-in options are inert when a strategy is provided; the good-since signal reaches the strategy once per connection, on the first event; exactly one failure is recorded when a read timeout is followed by the connection dropping (double-signal idempotence — verified to fail if the once() wrapper is removed); valid retry: values are forwarded; non-digit and negative values are ignored; over-cap values are clamped to 3600000; and a behavioral pin that a valid retry: governs the next reconnect delay with the built-in strategy (a previously untested path).
• Full suite: 98 passing, standard clean. Mutation check: the seam tests fail against the unmodified library except the negative cases, which are vacuous without the feature.

---

Note

Overview
Adds a retryDelayStrategy init option so callers can plug in custom reconnection timing via nextRetryDelay, setGoodSince, and setBaseDelay, fully replacing the built-in initialRetryDelayMillis / backoff / jitter options when set. EventSource.supportedOptions and the README document the new seam.

SSE retry: parsing is tightened: values must be all ASCII digits (rejecting negatives and parseInt-style partial matches), and accepted delays are capped at one hour before updating reconnectInterval and forwarding to the strategy.

Tests cover custom strategy behavior, retry: validation/capping, and built-in delay after a valid retry: field.

^{Reviewed by Cursor Bugbot for commit 7918fc5. Bugbot is set up for automated code reviews on this repo. Configure here.}