[two-tiger]

Summary

The project is renamed from Tandem to JevDuo (lowercase jevduo for anything machine-readable, JevDuo for display).

This PR is a mechanical rename only. It does not change behavior, the protocol, or the on-disk result format. It contains two commits:

1. refactor: rename tandem to jevduo (package, CLI, session class, paths)
2. docs: sync executable commands and imports with the jevduo rename; add upgrade note

Changes

Code and packaging

• src/tandem → src/jevduo, src/tandem_eval → src/jevduo_eval (git mv; 76 of 77 files are detected as renames)
• Distribution name jevduo; CLI entry points jevduo / jevduo-eval (the old tandem / tandem-eval commands are removed, no aliases)
• TandemSession → JevDuoSession
• rjob job name prefix tandem- → jevduo-
• pyproject.toml: the package-data key and the ruff per-file-ignores paths follow the new package path
• infra/config.example.json: repo_root / deploy_root / results_root
• infra/worker.sh: entry module jevduo_eval, docker temp dir prefix
• CI: tandem check → jevduo check
• Tests: imports, monkeypatch string paths, path / job-name assertions, and the regex in tests/demo/test_site.py that guards the "demo must not import the harness" boundary

Docs (executable content only)

• README / AGENTS.md / docs/*: commands, import examples, directory maps and the jevduo-<run-id> job name now match the renamed entry points, so the documented steps run as written
• README: new section "从旧名称 tandem 升级" (upgrade from the old name)

Compatibility impact (please review)

• Breaking: import tandem / import tandem_eval and the tandem / tandem-eval commands no longer exist. Existing installs need:

  pip uninstall -y tandem        # uv: uv pip uninstall tandem
  pip install -e '.[dev]'
  jevduo check

• rjob names become jevduo-<run_id>.
• The dev machine's infra/config.json (gitignored) needs eval.repo_root / eval.deploy_root updated by hand. The old tandem-deploy copy on the shared volume is left untouched and should be removed once no run uses it.

Intentionally not in this PR

• Env vars TANDEM_EXPERIMENT_ROOT / TANDEM_CONFIG_FILE (follow-up PR, with a fallback to the old names)
• harness.name, f"tandem {__version__}" in episode.json, the version number, and the schema $id / $ref / title (follow-up PR; wording and identifiers only, no protocol_version bump)
• Brand wording in README / AGENTS.md titles and intro sentences, and the DEVIATIONS.md history (wording PR)
• Shared-volume paths owned elsewhere: tandem_api_pool, tandem_jev_images, tandem_osworld21_*, and log.results_root (it points at the existing historical results directory)
• The real-machine capture in tests/eval/rjob_fakes.py (kept verbatim)

Verification

• ruff check src tests demo: pass
• pytest: 657 passed, 1 skipped (the skip is the existing optional-Chromium test, unavailable on my machine)
• jevduo check: 17 protocol samples pass
• Offline wheel build: the wheel contains the 3 schemas and 17 samples, entry points are jevduo / jevduo-eval, and there are no tandem entries (confirms the package-data key was updated correctly)
• After reinstalling the editable install, the entry points work and import tandem raises ModuleNotFoundError
• The README commands and the library-usage imports were executed for real: jevduo check, jevduo --version, jevduo-eval rjob run --help, jevduo-eval entry --help, and the build_default / SessionConfig / JevDuoSession imports
• The demo boundary guard regex was checked with positive and negative samples: from jevduo import x and import jevduo_eval match, from tandem import x does not

Cluster mock smoke

jevduo-eval rjob run --benchmark mock --limit 2

• run id rename-mock-20261105, rjob jevduo-rename-mock-20261105, code version 905c33b (no -dirty)
• Result: platform status Succeeded, rjob run returned succeeded, worker log {"done": 1, "scored": 1}
• Results: /mnt/shared-storage-gpfs2/evobox-share-gpfs2/chenxinquan/tandem/rjob_logs/rename-mock-20261105
• total is 1 because the mock manifest contains a single task (mock-000); --limit 2 did not drop anything
• Used the new jevduo-deploy as deploy_root; the old tandem-deploy copy was not touched
• The rjob no longer appears in the full rjob list (259 jobs listed, 0 matches). rjob run does not print the deleted flag and rjob get cannot distinguish "deleted" from "never existed", so this is indirect evidence of cleanup
• Scope: covers sync, submit, worker entry, result archiving and cleanup on the mock benchmark. It does not cover the real-benchmark path (shared env.sh, docker, VM; OSW1 / OSW2 / OSW2.1)
• Expected leftover: harness.name in the archived summary.json is still tandem (follow-up PR)

CI note

test (3.12) failed once on tests/eval/test_browser_controls.py::test_real_cdp_discovers_click_handlers_without_promoting_plain_text (a real-Chromium test timing out waiting for document.readyState). This PR only changes one import line in that file. The same test also failed on debug/gui-observation-closure and passed on re-run; it passed on re-run here too. The root cause was not investigated; hardening it should be a separate change.

Notes for reviewers

• Please view the diff with rename detection (git diff -M30% or GitHub's rename view). src/jevduo/__init__.py is only 3 lines with 49% similarity, so it shows as an add/delete at the default 50% threshold.
• tools/rank_probe.py:24 imports jevduo.candidates, which does not exist (the module is jevduo.gui.candidates). The old import was equally broken on the baseline, so it is not a regression of this rename. I will fix it in a separate small PR.