[adamlogic]

Summary

queueTimeFromHeaders treated every x-request-start value as whole milliseconds after stripping non-digits. Platforms that send fractional seconds with extra precision (NGINX) or whole nanoseconds (Render) reported 0 or NaN queue time, so queue-time autoscaling silently no-op'd.

This brings the Node parser in line with judoscale-ruby so those header variants produce a real queue time. Fixes #123.

Design decisions

Unit detection uses the same year-2000 magnitude cutoffs as the Ruby adapter, rather than inferring units from digit count or decimal places. That keeps both adapters agreeing on the same header, including t= prefixes and 2–6 decimal fractional seconds.

---

Note

Medium Risk
Changes queue-time calculation used by Express/Fastify/Next autoscaling metrics; behavior shifts for non-millisecond headers but is intended to fix silent misreporting rather than alter Heroku-ms paths.

Overview
Fixes x-request-start parsing so request queue time is computed correctly across Heroku, NGINX, and Render-style headers instead of often returning 0 or NaN.

queueTimeFromHeaders no longer strips non-digits and assumes milliseconds. It adds startedAtMs, which parses numeric values (including t= prefixes and fractional seconds), infers seconds vs ms/µs/ns using year-2000 magnitude cutoffs aligned with the Ruby adapter, and returns rounded Math.max(0, now - startedAt).

Tests cover NGINX fractional precision (2–6 decimals), microsecond, and nanosecond header values.

^{Reviewed by Cursor Bugbot for commit d882544. Bugbot is set up for automated code reviews on this repo. Configure here.}