test_harvest.sh §7 flakes intermittently with git plumbing errors ("could not parse HEAD", "could not read <sha>")

[moodmosaic]

Summary

tests/test_harvest.sh §7 ("Behavioural: harvest.sh --dry survives >20 commits") fails intermittently in ./tests/test.sh --unit with git plumbing errors that don't reproduce on a single-test re-run:

=== 7. Behavioural: harvest.sh --dry survives >20 commits ===
fatal: could not parse HEAD

or:

=== 7. Behavioural: harvest.sh --dry survives >20 commits ===
error: Could not read c8329c769513da7943f88d476e57ffbc63a02ebc
error: could not parse commit c8329c76...
fatal: the remote end hung up unexpectedly

Re-running ./tests/test.sh --unit (or ./tests/test_harvest.sh in isolation) passes 20/20. Five consecutive runs after a failure have all been clean.

Impact

Pre-merge gating: a CI run that hits this looks like a real regression when it isn't. Currently mitigated by re-running, but that erodes trust in the suite.

Likely root cause

§7 is by far the most I/O-heavy unit test: 100 --allow-empty commits with ~1 KiB subject padding, in a scratch repo under $TMPDIR, then git push to a scratch bare under /tmp.

Both error patterns are git's "I expected this ref/object to be readable and it wasn't" signature, which lines up with one of:

1. Auto-GC race. 100 loose objects approaches git's gc.auto threshold (default 6700 loose objects, but the trigger also depends on pack-refs timing). An auto-GC kicked off mid-§7 by an earlier test in the same suite could pack-then-prune under the test's feet.
2. Host config interference. test_harvest.sh does not export GIT_CONFIG_GLOBAL=/dev/null / GIT_CONFIG_SYSTEM=/dev/null (the isolation we added to test_launch.sh and test_session_end_push.sh in Pin Codex CLI, fix signing, isolate test gitconfig #91), so any host-side core.hooksPath, gc.*, maintenance.*, or pack.* config leaks in.
3. Filesystem hiccup on /tmp. Less actionable, but plausible on tmpfs/btrfs under load.

Proposed mitigations (not a fix yet — just lower flake rate)

Two cheap, additive changes worth trying first:

--- a/tests/test_harvest.sh
+++ b/tests/test_harvest.sh
@@ -1,5 +1,9 @@
 #!/bin/bash
 set -euo pipefail
+
+# Isolate from host gitconfig (gc.*, maintenance.*, hooks, etc.).
+export GIT_CONFIG_GLOBAL=/dev/null
+export GIT_CONFIG_SYSTEM=/dev/null

and inside §7's scratch repos:

@@ -198,6 +198,7 @@
 git clone -q "$HARVEST_BARE" "$HARVEST_AGENT"
 cd "$HARVEST_AGENT"
+git config gc.auto 0
 git checkout -q agent-work

If the flake survives both, the next step is GIT_TRACE=1 ./tests/test_harvest.sh on a failing run to see what git was actually doing when it lost the HEAD ref.

Repro

Not consistently reproducible. Observed during runs of ./tests/test.sh --unit on the master branch and on fix/codex-at-capacity-retriable (PR for #85), with two distinct error wordings on consecutive failures. Five clean re-runs in a row immediately after both failures.

Version

0.20.12 (and 0.20.13-pending on the codex-at-capacity branch).