Close ISSUE-LOCAL-01M5ZEVF7HQ8SFK9F8A9WSREEP and record the row outcome

.agents/issues/CI-AGENT-RECORD-DISK/ISSUE-LOCAL-01M5ZEVF7HQ8SFK9F8A9WSREEP.md

@@ -1,19 +1,19 @@
 ID: ISSUE-LOCAL-01M5ZEVF7HQ8SFK9F8A9WSREEP
 Title: agent-record flaky: strix disk-headroom exhausts on shared runner
 Row: CI-AGENT-RECORD-DISK
-State: OPEN
+State: CLOSED
 Kind: bug
 GitHub: -
 Mirror: PENDING
 Availability: FULL
 Created: 2026-10-27
-Updated: 2026-10-27
-Closed: -
+Updated: 2026-10-28
+Closed: 2026-10-28
 
 ## Problem
 
 The agent-record CI job fails intermittently on main (runs 40521291883, 40452744901) and reproduced on the #3513 rerun (job 122404172875): 49 tests in tests.tools.test_strix_vllm_oracle fail with ValueError('disk headroom exhausted') from tools/bench/strix_vllm_oracle/worker.py:257, plus CacheDropError (cache-drop proof retained 6291456 resident bytes). Environmental shared-runner pressure, not owned by a product row. Spec: .agents/specs/ci-agent-record-disk.md. Fix: free runner disk before the tools suite in .github/workflows/ci.yml.
 
 ## Resolution
 
--
+Landed 2026-10-28 as f484c76ad (PR #3515). The wall was not runner disk pressure on / but the /tmp tmpfs of the ubuntu-26.04 image (actions/runner-images#14777): the strix worker's headroom() measures disk_usage of self.local and self.output, both under /tmp (worker.py:226,256), so the 20 GiB floor was unmeetable on 26.04 runners regardless of how empty / was. Proof the first attempt was insufficient: run 40586147690 job 122477980562 showed 113 GiB free on /dev/root and still failed 49+2 strix tests; build-test-cpu (122477933406) and sanitize-cpu (122477933291) failed identically in the same run. Fix: mask tmp.mount and lazy-unmount /tmp as the first step of the six ubuntu-latest compile/ctest lanes. Verified green: run 40637909804 -- agent-record and build-test-vulkan passed with df -h / /tmp showing /tmp on /dev/root at 88 GiB free.

.agents/specs/ci-agent-record-disk.md

@@ -149,6 +149,33 @@ locally.
   workaround (mask `tmp.mount`, lazy-unmount `/tmp` as the first step) is
   the one this change lands, and runner-images#14814 is the image-side fix.
 
+## Outcome
+
+Landed as f484c76ad (PR #3515, merged 2026-10-28T11:05Z). What settled it:
+
+- The first hypothesis (shared-runner disk pressure on `/`) was **rejected
+  by evidence from its own fix**: PR #3515's first commit freed `/` to
+  113 GiB and the job still failed 49+2 strix tests. The worker measures
+  `/tmp` filesystems, not `/`.
+- The root cause is the ubuntu-26.04 image's `/tmp` tmpfs
+  (actions/runner-images#14777), which makes the worker's 20 GiB floor
+  unmeetable by construction; the flakiness is runner-image assignment
+  during the 24.04→26.04 rollout.
+- Defaults: the mask+lazy-unmount step lands FIRST in six ubuntu-latest
+  compile/ctest lanes (`agent-record`, `build-test-cpu`, `sanitize-cpu`,
+  `build-test-vulkan`, `build-newest-gcc`, `cuda-fat-build`), and only
+  those — the trivial record/document lanes do not write `/tmp`, and the
+  pinned `ubuntu-24.04-arm` lanes keep `/tmp` on the root disk. The
+  agent-record reclamation step stays because `/tmp` returning to the root
+  disk makes the freed space reachable by the suites and compilers. The
+  worker's 20 GiB floor is untouched: a safety property of the oracle
+  runner is not weakened to green a gate.
+- Verified green on run 40637909804: `agent-record` and
+  `build-test-vulkan` passed with `df -h / /tmp` showing `/tmp` on
+  `/dev/root` at 88 GiB free; the long ctest lanes and the pre-existing
+  windows-msvc flint.cpp warning-as-error break are recorded separately
+  and do not belong to this row.
+
 ## Stop conditions
 
 - If GitHub's runner image changes so that the reclaimed paths no longer