fix(pacing): anchor intervals to transport starts

[Wibias]

Summary

• make provider request pacing measure the interval at the actual outbound transport start instead of before an awaited continuation;
• replace the wall-clock-sensitive concurrency assertion with an injected-clock regression that reproduces delayed continuation starts deterministically.

Verification

• RED phase: test-only commit intentionally reproduces the current scheduling gap in CI before the implementation change.
• Final focused/full checks will be recorded after the production fix lands.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed. No docs change is required because this restores the documented/configured pacing guarantee without changing configuration.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. This change does not touch credentials, authentication, or secret handling.

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 32665ad3-a888-4837-82b3-9787a1ddf3ed

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[lidge-jun]

Reviewed as part of the Wave 5B train and held, not rejected — the direction is right and
three of the five things this needed are already here.

Present and verified in the diff:

• intervals anchored to transport start rather than queue entry
• a cancelled waiter does not consume a slot (aborted queued requests leave immediately and never consume a start)
• deterministic timing via an injected clock rather than wall time

Two are missing, and they are the ones that decide whether the pacing clock is actually
correct under load:

1. Retry does not double-advance the clock. A retried request that re-enters the pacer
   must not count as a second transport start, or the effective interval silently doubles
   under exactly the conditions that trigger retries.
2. Per-account isolation. account appears zero times in this diff. If one account's
   pacing can delay another's, the fix trades a burst problem for a starvation problem.

One smaller thing: the description still says "Final focused/full checks will be recorded
after the production fix lands," but the production fix is already in the diff — 56 lines in
request-pacing.ts and 72 in fetch-helpers.ts, including the
waitForProviderRequestSlot → runWithProviderRequestSlot signature change. Worth refreshing
so a reviewer is not told to expect something that already happened.

Happy to see this land once those two tests exist. The PR is still draft, so nothing is
blocked by this note.