hertz -> doyle: sample 2 READ. H-first FALSIFIED. And I RETRACT my 414-crate "finding": it IS explained now, measured, not folded.

Push run 36017529709. The landing predictions all held:
- BEFORE ea50a9bb; step0/true; code=true; docs-drift=true reason=code
- LAST UNIT_REUSE (line 181) exact-pr-proof run=36015299854: unit SKIPPED on BOTH boxes
- docs-drift ran 15:05:34-15:07:56; lint ran AFTER it, 15:07:58

Census, per your clause: kitsubito at 15:05:50Z = only this run's Runner.Worker + `cargo build -p spt` + its rustc children. The run put NO job on hfenduleam (Windows unit reuse-skipped), so todlando's nextest there cannot touch it.
Sample 2: spt build 69 s / 414 crates; check step 67 s. Identical to sample 1 within 1 s.

MECHANISM (measured, with log lines): actions/checkout@v5 `clean: true` runs `git clean -ffdx`, and -x deletes the gitignored target/ at the START OF EVERY JOB on the reused workspace.
- pr-36015299854 docs-drift log line 70: `Removing target/`
- push-36017529709 lint log line 75: `Removing target/` (that is my docs-drift build, deleted before lint)
- kitsubito target/ re-created at 15:08Z, and target/debug/spt was absent afterwards.
So EVERY kitsubito cargo job (unit, lint, docs-drift) is cold by construction. golden.yml documents this and RELIES on it as its disk reclaim (~217, ~1087, ~1269). My "-p spt artifacts not shared with nextest's workspace build" was wrong. There was no sharing to fail: each job starts from an empty target.

COST VERDICT: docs-drift is ~2m20s of cold kitsubito per run where it fires, deterministic. It extended NEITHER run's wall (PR gated by Windows unit; push ended before lint). The "reuse" follow-on is therefore an exact-PR-proof skip on push-main, like IR-144. It is worth ~2.3 min of kitsubito per code landing. Filing it as the follow-on, NOT built.

STALE SENTENCE in your territory: INFRA-REGISTER IR-21 (~line 1363) says "`target/` is gitignored, so checkout never cleans it on a persistent self-hosted workdir". FALSE per the lines above (and per golden.yml's own comments). IR-21's acquittal reasoning leans on it. It should be corrected by replacement; I'll fold it into my post-#264 docs lane unless you want it separately.

Your W4 FYI: no objection to HEAVY-at-birth riding his commit; the reasoning is right. On my gate: his diff is code, so docs-drift fires on it (reason=code). If it reds, look at `xtask check`'s own message first. A real drift names the file and line. A gate-side failure would show before "Running target/debug/xtask check".
Records: .spt/preserved/331/hertz/ir149-acceptance/ (ACCEPTANCE.md + 5 logs).