[aviggiano]

Summary

Removes a causal claim I merged in #305 that measurement disproves.

What was wrong

#305's doc comment attributed four sandbox deaths at stateful-invariant-setup to a cgroup OOM inside git add. Reviewers of the follow-up refuted it, and I reproduced the disproof rather than take their word for it:

$ head -c 155000000 /dev/zero > big{1,2,3,4}.bin   # 620 MB across four blobs
$ /usr/bin/time -v git add -A
        Maximum resident set size (kbytes): 155812

git add peak RSS is per-file, not cumulative — 620 MB across four blobs peaks the same as one — and a single blob above core.bigFileThreshold streams at a few megabytes. Against memoryMiB: 32_768 / memoryLimitMiB: 65_536 (packages/modal/src/defaults.ts:28-32), staging the observed 575 MB workspace costs well under 1% of the memory request.

It has since been settled from the other direction too: R47 attempt 1 recorded worker_exit_code: 1 — the first exit code ever captured for one of these deaths, thanks to #303. An OOM kill is 137. Exit 1 is a deliberate process exit, so no memory theory explains it.

Why this is worth its own PR

Leaving a disproven causal story in the source is the exact defect I spent two PRs fixing today in the prompts: text that contradicts what the code and environment actually do, discoverable only by someone burning a node on it. A future maintainer reading that comment would have started from a dead end I already walked.

The comment now carries the measurement, the sandbox limits, and a plain statement that the cause is unidentified — plus an explicit warning not to read it as a diagnosis.

What is unchanged

The exclusion itself, and its justification: harness-generated corpus is not authored content, nothing downstream consumes it from a patch, and MAX_PATCH_BYTES meant no corpus above 16 MB ever survived a handoff. It simply is not the fix for these deaths.

The size-ceiling follow-up (#306) is closed, for the same reason plus three independent defects review found in it.

Validation

Comment only — no behaviour change. workspace-handoff 16/16, typecheck, eslint, prettier clean.

Refs #304, #307

Greptile Summary

This PR replaces an unsupported OOM diagnosis in the workspace-handoff documentation with measured evidence identifying oversized staged diffs and ENOBUFS as the observed failure.

• Documents the ordering between runGit’s 32 MB buffer limit and the 16 MB patch-size check.
• Records why generated corpus exclusions are operationally necessary.
• Clarifies that measured git add memory use does not support the previous OOM claim.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
packages/runtime/src/workspace-handoff.ts	Updates an operational comment explaining the workspace-patch exclusion and observed failure without changing runtime behavior.

_{Reviews (2): Last reviewed commit: "docs(runtime): state the verified cause ..." | Re-trigger Greptile}

Context used:

• Knowledge Base — Runtime: Run Orchestration Lifecycle