Measure the effect of no longer requiring pull-request rebases #2547

Goal

Find out whether making pull-request rebases optional (issue #2507) reduced force-pushes, pull-request CI runner time, and rebase-fallout commits. Compare a two-week window under the new policy with the baseline that #2507 recorded for 1 to 13 October 2026.

Background

Issue #2507 stopped requiring a rebase onto develop before merge. The ADR was accepted in spec PR #2508, merged 2026-10-14 06:42 UTC. The maintainer merge tool was changed in PR #2526, merged 2026-10-14 12:18 UTC, to refresh a pull request's stale recorded base instead of asking for a rebase. The goal was to cut CI reruns, review noise, and serialized merging.

The #2507 specification planned this comparison as its task T7 and acceptance criterion AC9. It needs at least two weeks of merges under the new policy. The maintainer decided on 2026-10-15 to move it into this issue so #2507 could close once its implementation and manual verification were complete.

Issue #2507 recorded the baseline, with reproduction commands, in baseline-evidence.md. That file moves to docs/issues/closed/2507-stop-requiring-pull-request-rebases/ when #2507 is archived. The baseline has these sections:

• section 4: pull-request volume and queue;
• section 5: force-pushes per merged PR;
• section 6: CI cost of pull-request runs;
• section 7: rebase-fallout commits.

The early merges are already mixed. Of the 18 merges from #2526 to #2543 (2026-10-14 12:18 to 2026-10-15 16:33 UTC), nine were merged while behind develop without a rebase, and six of those relied on the merge tool's refresh. Recommended rebases are still common, though: for example, #2530, #2538, and #2430 rebased clean branches. See #2507 manual-verification-evidence.md, section "Merges Under the New Policy".

Scope

In Scope

• Measure the window from 2026-10-15 00:00 UTC to 2026-10-28 23:59 UTC (14 days) with the Stop requiring pull-request rebases and reconcile repository rulesets #2507 baseline commands, and normalize per merged PR and per day so the 13-day baseline stays comparable. This specification fixes how each measure selects its data, because the baseline sections do not all say how they selected theirs:
  • force-pushes per merged PR and the share of merged PRs force-pushed at least once. The population is the pull requests merged in the window (merged_at), with every force-push of each of them, including events before 2026-10-15: baseline section 5 selects pull requests by merge date and counts each one's whole force-push timeline. The classification below uses the same population;
  • estimated PR CI runner-hours per workflow, sampled as in the baseline. The runs are the pull_request workflow runs created within the window, not the runs of the pull requests merged in it, so runs of pull requests still open count too. Baseline section 6 reports 899 runs for 1 to 13 October and names no date field; creation date is this specification's reading of it;
  • rebase-fallout commits, compared with a baseline recounted for 1 to 13 October only: section 7 of the baseline lists 23 commits "since September 2026", of which 13 have author dates from 1 to 13 October, so do not divide the whole list by 13 days or by the 67 October merges. A commit is in the window when its author date is, as in that 13-of-23 recount. The command in section 7 filters with --since, which is the committer date, and prints author dates;
  • pull-request volume and the daily open-PR queue, for context, as in baseline section 4.
• Classify every force-push of those pull requests with the method of Stop requiring pull-request rebases and reconcile repository rulesets #2507 manual-verification section V4, which compares where the push put the head's base, not only where the new head sits. The develop tip of its time is the last first-parent develop commit whose pull request has a merged_at before the push. A rewrite of a head that was already on the tip must not be counted as a rebase:
  • rebase: the push moved the base, that is, git merge-base <tip> <beforeCommit> differs from git merge-base <tip> <afterCommit>. It is required when the pushed-away head conflicted with the tip (git merge-tree --write-tree), and recommended when it merged cleanly;
  • not a rebase: the push left the base unchanged. Since PR docs(agents): allow rewriting an unmerged pull request's history with safeguards #2531 merged (merged_at 2026-10-14 16:09 UTC), AGENTS.md allows other rewrites of an unmerged pull request's history, such as splitting or reordering commits, which the baseline could not contain: baseline section 5 counts every force-push as a rebase because history rewrites were otherwise forbidden.
    Report the total force-pushes per merged PR, which is the baseline's measure, and the rebase-only figure, which excludes the second class.
• Count merges made while behind develop whose head was pushed before develop reached M^1, the develop tip the merge was built on, and not pushed again before the merge (the recorded-base refresh path). The cutoff is M^1, not the Git merge base, which a head always postdates. As in Stop requiring pull-request rebases and reconcile repository rulesets #2507 manual-verification section V1, the count is inferred, because the merge tool's output is not captured.
• Record the comparison and a short conclusion in issue-local post-change-evidence.md.

Out of Scope

• Changing the rebase policy, the ADR, the skills, or the merge tool. If the results argue for a change, propose it in a new issue.
• Building scripts or tooling for the measurement. Ad-hoc gh and git commands are enough. Record them verbatim.
• Measuring the CI gate work of Require one always-reporting CI gate decided by a Rust tool #2525. If Require one always-reporting CI gate decided by a Rust tool #2525 changes PR workflows during the window, record which changes landed and when, so they are not mistaken for an effect of the rebase policy.

Plan

• T1: do not start before 2026-10-29; record any Require one always-reporting CI gate decided by a Rust tool #2525 workflow change in the window.
• T2: collect the measurements with the Stop requiring pull-request rebases and reconcile repository rulesets #2507 baseline commands (sections 4 to 7) plus the V1 and V4 classifications, recounting the rebase-fallout baseline for 1 to 13 October only.
• T3: compare with the baseline, normalized per merged PR and per day, and conclude; propose any policy change in a new issue.

The specification is added under docs/issues/open/ with the #2507 close-out pull request.