Round 4 (`CODEX-REVIEW-L18-JOINT-ROUND-FINAL-4.md`) closed finding 1 and raised two P2s. Both are now fixed.
4
5
1. **The relocation probe dropped `Staged`. Fixed in 3320b9ce on this branch.** `RelocationWakeInputs.Select` (`internal/host/relocationwake.go`) now forwards `Staged: l.Staged`. Every path through `Select` shares that one value.
6
7
The builder found no other place that drops it:
8
- the serving wake after completion uses `claimowners.go`, which already forwards it;
9
- `ColdHandover` builds no launch inputs;
10
- `relocationinstall.go` and `relocationassembly.go` construct only `DBPath`.
11
12
The test `TestTheRelocationProbeStartNamesTheBuildItHashed` drives the real `activator.ClaimRuntime`. It fails without the fix. It copies `claimedBuild`'s condition rather than calling it, because that function is unexported.
13
2. **The model's dirty return ignored target admission. Fixed in `region-ha-two-box-build` at 3b895b42.** Three places now require the claim to record the target's admission of this move, mirroring `movesinstall.go`'s `targetAdmitted`:
14
- `filesHome`;
15
- the INV-30 oracle `servableRegions`;
16
- the install park.
17
18
The new witness `wDirtyReturnWithoutAdmission` covers this, and the new mutation `dirty-return-needs-admission` is killed by INV-30. `TestFilesWitnessesStuckWithoutINV30` now skips the new witness: that scenario has no liveness consequence, so INV-30 is its only oracle. The unmutated 3000-seed run had no failures.
19
20
**Judge:**
21
1. Are both round-4 items closed?
22
2. Is the `TestFilesWitnessesStuckWithoutINV30` exclusion sound?
23
3. Anything new? Only report what these two diffs introduce. Do not re-review settled rulings or the main merge again.
amadan