---
name: a-push-to-main-runs-the-full-thin-lane-on-both-runners
description: "SUPERSEDED 2026-09-24 by IR-147 (0854bbe8): a docs-only push to main now classifies over the whole push range and SKIPS lint/unit/docs-drift, same as a PR. The standing rule stays: read the PUSH run's job list, never infer it from the PR run."
metadata:
  node_type: memory
  type: project
  originSessionId: e79a9f87-6f3c-4459-a887-818024d2da0f
  modified: 2026-09-24T15:45:20.701Z
---

**Current (measured 2026-09-24, hertz):** since IR-147 (`0854bbe8`, "classify docs-only
pushes to main over the whole push range"), the push run classifies too. Push run
36022121680 at `07d04ba5` (core PR #266, AGENTS.md + INFRA-REGISTER only) logged
`CLASSIFY_PROBE connected_at=deepen50 shallow=true`, `code=false reason=docs-only`,
`docs-drift=false reason=no-drift-input`; unit, lint and docs-drift were SKIPPED. A push whose
range has code still runs the full lane on both runners. So a docs-only push no longer loads
hfenduleam, but ONLY if the WHOLE push range is docs-only.

**Superseded premise (2026-09-09, doyle):** ci.yml `changes.classify` (line 49 at `88701366`)
had `if event_name != pull_request -> code=true`, so ff pushes of docs-only PRs #206/#207 ran
40+ min Windows unit jobs through hertz's IR-85 timing window after doyle had promised
"kitsubito only". That line is gone. Do not cite it.

**Why the rule survives:** a push and a PR are two events with two classifications, and a
push range can span more than the PR (several landings, a base move). Evidence about one event
is not evidence about the other. Concurrency on main is `cancel-in-progress: false`, so
queued pushes run back to back.

**How to apply:** before promising a runner is untouched, `gh run list --commit <sha>` and
read the job list of the push you are about to fire, or check that the whole push range is
docs-only. Batch docs lanes into one push during a timing window anyway. Related:
[[open-the-subindex-before-running-a-gate]], [[a-stale-target-makes-a-check-pass-harder-not-fail]].