fix(test/acp): make TestStartLongSocketPathUsesShortSocketName robust to TMPDIR-length variance

[scarson]

Summary

TestStartLongSocketPathUsesShortSocketName in internal/runtime/acp
flakes on macOS under make test-fast-parallel. The bead-author
hypothesized "socket name collision under parallelism," but the real
cause is purely arithmetic: the test's valid-path window is ~8 bytes
wide, but the loop iterates in 10-byte steps and so can skip over the
window entirely depending on len(root).

root is built from os.MkdirTemp(\"\", \"gc-acp-sock-\"), whose random
suffix is itoa(rand.Uint32()) — a variable-length string of 1–10
digits. On default macOS ($TMPDIR = 49 bytes), most suffix lengths
land root at 70-71 (works), but ~2.3% of rand.Uint32 values
stringify to 7-8 digits and produce root = 68 or 69, which lies in
a gap where no i ∈ [1, 32] satisfies both legacy > 108 and
short < 108. The test then t.Fatals at the very start (Elapsed=0.02s
matching the bead's observation).

Confirmed by reproduction simulating multiple root lengths — gaps at
roots 68, 69, 78, 79, ≥85 — and by direct repro with an elongated
TMPDIR that fast-failed before the fix and Skips gracefully after.

Fix

• Step by a single byte (strings.Repeat(\"x\", i)) so the loop can
  always land in the valid window for any reasonable len(root).
• Compare against 104 (macOS sun_path) rather than 108 (Linux), so
  the constructed path actually binds on both platforms. (At 10-byte
  steps the slack was wide enough that this never bit; at 1-byte steps
  it could.)
• If TMPDIR is so long that no valid path can be constructed at all
  (root ≥ 85), Skip with a diagnostic. That state is the
  fundamental constraint — not a bug — and shouldn't appear as a Fail.

Why this candidate

The bead description suggested "use t.TempDir() / per-test unique
socket name" as a likely fix. Neither matches the actual root cause:

• t.TempDir() produces longer paths (includes the test name), which
  pushes root further into the impossible region — would make the
  flake worse, not better.
• "Per-test unique socket name" doesn't apply: the test's name = \"control-dispatcher\" is intentional (it's exercising the path-length
  fallback for that exact name) and the socket lives in a unique
  MkdirTemp directory per run, so cross-test collision was never the
  issue.

Adjusting the loop's step size and the platform check is the smallest
change that actually fixes the failure mode.

Repro

Before:

$ TMPDIR=/var/folders/.../T/extra-long-prefix-12345 \
    go test -count=1 -run TestStartLong -v ./internal/runtime/acp/
--- FAIL: TestStartLongSocketPathUsesShortSocketName (0.01s)
    acp_test.go:741: failed to construct path where legacy socket is
    too long but short socket fits

After:

$ TMPDIR=/var/folders/.../T/extra-long-prefix-12345 \
    go test -count=1 -run TestStartLong -v ./internal/runtime/acp/
--- SKIP: TestStartLongSocketPathUsesShortSocketName (0.05s)
    cannot construct path where legacy socket exceeds 104 bytes but
    short socket fits; len(root)=95 (likely TMPDIR=\"...\" too long)

$ GC_FAST_UNIT=1 go test -count=20 -p=8 -run TestStartLong \
    ./internal/runtime/acp/
ok  github.com/gastownhall/gascity/internal/runtime/acp  1.200s

Full package: go test -count=1 -p=4 ./internal/runtime/acp/ → ok 3.443s.

Trail

• PR test: skip 127.0.0.0/8 alias bind on darwin (1 of 2 pre-existing macOS flakes) #2063 (fix/macos-test-flakes) fixed one of the two reported macOS
  pre-commit flakes and identified this one (along with a sibling) as
  the actual underlying issues.
• Bead: ga-urt (gascity rig).
• Sibling beads from the same investigation: ga-un5, ga-kkn.

Testing

• go test -count=1 ./internal/runtime/acp/... (the affected
  package)
• go vet ./internal/runtime/acp/
• -count=20 -p=8 stress, plus a forced-failure reproduction
  that now correctly Skips
• make check — the pre-commit hook surfaced one unrelated
  pre-existing flake (TestResolveDoltConnectionTargetManagedCity_EnvOverride
  in internal/beads/contract) that is the exact flake fixed by
  in-flight PR test: skip 127.0.0.0/8 alias bind on darwin (1 of 2 pre-existing macOS flakes) #2063. That fix is not yet on origin/main where
  this branch is based; only the acp test is changed here and the
  failing test is in a different package.

Checklist

• Linked an issue (bead ga-urt) and the trail PR (test: skip 127.0.0.0/8 alias bind on darwin (1 of 2 pre-existing macOS flakes) #2063); no
  upstream GitHub issue per bead guidance
• Added/updated tests for behavior changes (the test under test is
  itself the change)
• Updated docs for user-facing changes (n/a — internal test fix)
• Called out breaking changes or migration notes (n/a)

🤖 Generated with Claude Code