Fix #364: [milestone Milestone 6 ] Add examples/divergence_demo.py: load two pre-recorded dumps, invoke `replay_d...

[telleroutlook]

Fixes #364

Generated by claude-bot-go worker.

[telleroutlook]

Summary

Internally consistent diff: replay_diff correctly loads two dumps, rank-aligns action streams, and reports the first divergence; tests verify the asserted outputs and the restricted-unpickler round-trip. No blockers.

Severity

low

Blocking findings

1. train_replay/replay/replayer.py:158 replay_diff uses no instance state — every helper it calls (_load_bundle, _group_actions_by_rank, _first_divergence) is a @staticmethod, so the graph passed to EpochReplayer is unused for this operation. Could be a staticmethod or standalone function; current shape works but couples the API to an instance needlessly.
2. train_replay/replay/replayer.py:233 The length-mismatch branch (_first_divergence at the len(baseline) != len(candidate) check, which triggers _missing_record sentinel records) is not exercised by any test — tests only cover value divergence and byte-identical dumps, leaving the missing-record path unverified.

Verdict

✅ Approved

Merge risk

low

Audit

• model: opus
• effort: high
• tokens: input=0, output=0

{
  "approved": true,
  "severity": "low",
  "summary": "Internally consistent diff: replay_diff correctly loads two dumps, rank-aligns action streams, and reports the first divergence; tests verify the asserted outputs and the restricted-unpickler round-trip. No blockers.",
  "findings": [
    {
      "file": "train_replay/replay/replayer.py",
      "line": 158,
      "issue": "replay_diff uses no instance state — every helper it calls (_load_bundle, _group_actions_by_rank, _first_divergence) is a @staticmethod, so the graph passed to EpochReplayer is unused for this operation. Could be a staticmethod or standalone function; current shape works but couples the API to an instance needlessly.",
      "kind": "suggestion"
    },
    {
      "file": "train_replay/replay/replayer.py",
      "line": 233,
      "issue": "The length-mismatch branch (_first_divergence at the `len(baseline) != len(candidate)` check, which triggers _missing_record sentinel records) is not exercised by any test — tests only cover value divergence and byte-identical dumps, leaving the missing-record path unverified.",
      "kind": "suggestion"
    }
  ],
  "merge_risk": "low",
  "Model": "opus",
  "model": "claude-cli",
  "effort": "high",
  "tokens": {
    "input_tokens": 0,
    "output_tokens": 0
  },
  "same_verdict_count": 1
}

[telleroutlook]

Merge gate failed: go test ./... exited non-zero.

verify manifest: /srv/claude-bot/worktrees/WasmAgent_wasmagent-train-replay/.gate-1785127652652920058/.claude-bot/verify.yml
  [PASS] install (SLUG=$(basename "$PWD"); if timeout 8 bash -c "test -d /mnt/bigdisk/venv && touch /mnt/bigdisk/venv/.probe-$SLUG 2>/dev/null && rm -f /mnt/bigdisk/venv/.probe-$SLUG" 2>/dev/null; then VENV=/mnt/bigdisk/venv/$SLUG; echo "[venv] bigdisk: $VENV"; else VENV="$PWD/.venv-local"; echo "[venv] local fallback: $VENV"; fi; { [ -x "$VENV/bin/python" ] || python3 -m venv "$VENV"; } && rm -rf .venv && ln -sfn "$VENV" .venv && .venv/bin/pip install -e ".[dev]" -q --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu) took 4m54.534s
  [PASS] python-import (.venv/bin/python -c "import train_replay; import train_replay.agent; import train_replay.replay; import train_replay.recording") took 109ms
  [PASS] lint (.venv/bin/python -m ruff check train_replay tests) took 438ms
  [PASS] typecheck (.venv/bin/python -m mypy train_replay) took 3.705s
  [FAIL] test (.venv/bin/python -m pytest tests/ -x) took 705ms

Failing command output: