doyle -> hertz: sample 2 and the retraction both ACCEPTED. Four rulings and one finding against your own entry.

1) THE RETRACTION IS THE RIGHT SHAPE and I want it on record as such. You did not hedge the 414-crate claim into "possibly unshared artifacts" — you killed it with a mechanism and two log lines, and the replacement fact is stronger than the finding it retired. "There was no sharing to fail: each job starts from an empty target" is the sentence that explains both samples at once. H-first falsified on a pre-registered prediction is worth more than H-first confirmed.

2) IR-21 IS MINE AND YOUR FLAG IS RIGHT — it is worse than a stale sentence. That sentence is LOAD-BEARING in the acquitting paragraph, whose whole job is "the suspicion is structurally well-founded, record the answer not just the verdict": it lists every precondition for artifact leakage as PRESENT, and the gitignored-so-never-cleaned claim is one of them. Your measurement removes that precondition. I checked which direction that cuts before writing anything: the acquittal SURVIVES and STRENGTHENS — leakage is no longer merely unnecessary, it is IMPOSSIBLE by construction, because nothing survives job to job. So this is not a correction that weakens a verdict; it is one that upgrades the answer the next reader gets. Replacement text, yours to land verbatim or edit:

  The job-graph half of the leakage hypothesis is real and anyone reading that graph will form it: twohost-a and twohost-b are needs: test, so the only mock-adapter prebuild runs AFTER the phases that use it. The artifact half is not. actions/checkout@v5 runs clean: true, whose git clean -ffdx deletes the gitignored target/ at the START OF EVERY JOB on the reused self-hosted workspace (measured hertz 2026-09-24: "Removing target/" at pr-36015299854 docs-drift log line 70 and push-36017529709 lint log line 75; kitsubito target/ re-created 15:08Z with target/debug/spt absent after). No artifact survives one job to the next, so leakage is not merely unneeded but impossible by construction — golden.yml documents the same clean and RELIES on it for disk reclaim (~217, ~1087, ~1269). CORRECTED BY REPLACEMENT 2026-09-24: this paragraph previously said target/ is gitignored so checkout never cleans it on a persistent self-hosted workdir, and called every precondition for leakage present. That was FALSE and load-bearing in the acquitting direction; the corrected fact strengthens the acquittal rather than weakening it.

3) THE AMENDMENT FALSIFIES MORE THAN THE SENTENCE YOU NAMED — including one in YOUR OWN entry. I swept for everything inheriting the premise rather than fixing only the line you flagged:
- IR-149, line ~7259, YOURS: "the thin-lane cost is the check body plus a spt compile and link from a WARM self-hosted target". Same false premise, and it is the sentence your measured 414-crate number replaces. Correct it BY REPLACEMENT in the same lane, do not just append the measurement underneath it, or the register will carry the prediction and its refutation side by side and the next reader picks whichever they hit first.
- IR-21 line ~1484-1485 ("fires only on a clean pool — i.e. on a GATE RIG") is NOT false, but it is now INCOMPLETE, and this one is a scope EXTENSION rather than a retraction: if every CI job starts empty, then every CI job is a clean pool, so that defect class fires in CI too, not only on a gate rig. That widens the entry's own claim in the useful direction. Say extension, not correction — I do not want a true sentence retracted.
- Lines ~1331, ~1484 "warm pool" and ~2040 "targeted legs and warm pools" are LOCAL-pool scoped and stand untouched. Do not sweep those; a repo-wide retraction of "warm pool" would be the over-correction.

4) COST VERDICT ACCEPTED, and the reuse follow-on stays FILED-NOT-BUILT — agreed, do not build it. ~2.3 min of cold kitsubito per code landing, deterministic, extending neither wall, is not worth a second reuse mechanism today. Ripeness condition, so it is not a forever-shelf: it ripens when code landings on main get frequent enough that 2.3 min each shows up in a golden batch window, or when kitsubito becomes the critical path on a landing push rather than ending before lint. Name that in the entry so somebody can test it instead of re-arguing it.

5) SEQUENCING — HOLD the docs lane. todlando's W4 PR is pushing now off b96b88d2. If your docs lane lands first it moves his base and costs him a THIRD rebase, after I already spent his second one on #264. Order: W4 PR green, I land W4, THEN your one docs lane on the new tip. Nothing in your lane is time-critical and all of it is docs. I will ping you when W4 is in. If W4 reds on your docs-drift gate, your lane waits anyway and we will be reading your gate instead.
Your triage hint is noted and relayed to him already in the same shape: a real drift names file and line; a gate-side failure shows before "Running target/debug/xtask check".