Improve latency calculation based on server timings and connect times

[carlosefr]

Improve accuracy of latency calculations by making use of additional "thinking time" sent by the server, and by using the observed TCP RTT as fine tuning.

[ask-bonk]

Review: Improve latency calculation based on server timings and connect times

Summary

This PR improves latency accuracy in the BandwidthEngine by:

1. Richer Server-Timing parsing — cfGetServerTime now matches cfReqDur/cfRequestDuration (case-insensitive, with variants) and falls back to summing cfSpeed* sub-metrics, replacing the previous single generic dur= match.
2. TCP RTT fine-tuning — a new serverTimeDelta is calibrated from fresh TCP connections (HTTP < 3) by comparing the derived server time (ttfb - connectTime) against the server-reported time, blended with an EMA (weight 0.75). This delta is threaded through MeasurementEngine, ParallelLatency, and persisted across sub-engines so calibration carries forward.
3. Config/doc cleanups — estimatedServerTime default changed 10 → 0; logAimApiUrl default repointed from aim.cloudflare.com/__log to speed.cloudflare.com/__results; matching README/AGENTS.md doc updates.
4. Idle labeling — idle (non-loaded) latency requests now tagged with during=idle query param.
5. Tests — new unit test covering the three cfGetServerTime branches; cfGetServerTime is now exported.

Overall the approach is reasonable and unit tests pass (101/101). The main issues are the unconditional debug console.log statements, which should not ship in a library, plus a few correctness edge cases worth confirming. Details inline.

[ask-bonk]

Review posted successfully to PR #127 with 7 inline comments. Here's a summary of what I reviewed and flagged:

What the PR changes

Improves latency accuracy in BandwidthEngine via three main mechanisms:

1. Richer Server-Timing parsing — cfGetServerTime now prefers cfReqDur/cfRequestDuration (case-insensitive variants) and falls back to summing cfSpeed* sub-metrics.
2. TCP RTT fine-tuning — a new serverTimeDelta, calibrated from fresh TCP connections (HTTP < 3) using an EMA (weight 0.75), threaded through MeasurementEngine and ParallelLatency and carried forward across sub-engines.
3. Config/doc cleanups — estimatedServerTime default 10 → 0, logAimApiUrl repointed to speed.cloudflare.com/__results, plus README/AGENTS.md updates and during=idle labeling.

Also adds a unit test for cfGetServerTime (all 101 unit tests pass locally).

Issues flagged inline

• High: Unconditional console.log debug statements (lines 469–504) ship to all library consumers on every request — inconsistent with the codebase's console.warn-only convention.
• Correctness: ping floor lowered from 1e-2 to 0, allowing 0ms latency data points that can skew percentile/jitter stats.
• Correctness/confirm: serverTimeDelta only propagates back via onFinished/onConnectionError — may be lost on other abort paths.
• Confirm intent: default endpoint change (aim.cloudflare.com/__log → relative /__results) and estimatedServerTime default change (impacts HTTP/3 where calibration is disabled).
• Minor: cfSpeed[a-zA-Z]* regex silently sums any future cfSpeed-prefixed metric; connectTime reused-connection semantics worth a clarifying comment.

The review was posted as COMMENT (not approve/request-changes), and no commits were pushed.

github run