Verification: PASS ✅

Ran: bash tests/run_all.sh on the chore/test-refactor branch — all 3 suites pass (unit_codegen, e2e_init_patch, e2e_lifecycle). Core behaviours verified manually:

• wtcraft init correctly copies llm-anti-patterns.md and presets/
• wtcraft init --patch-agent-files injects the new routing stub into CLAUDE.md
• gen-presets.py runs cleanly; generated presets match committed files

Findings (3 issues worth addressing before merge)

⚠️ 1. Language enforcement stub is user-hostile for international audiences

scripts/wtcraft injects this into every user's CLAUDE.md when they run wtcraft init --patch-agent-files:

CRITICAL: You must output YOUR ENTIRE RESPONSE in **Simplified Chinese (简体中文)** or English ONLY. 
Under absolute zero circumstances should you output any Hiragana, Katakana, or Hangul characters.
Any use of Japanese or Korean characters will cause an immediate system crash.

Problems:

1. wtcraft is a public OSS tool. A French, German, or Japanese developer running wtcraft init will have their CLAUDE.md silently locked to Chinese/English. This is a breaking change to their workflow.
2. "will cause an immediate system crash" is false — it's a Claude instruction, not a system property. Misleading wording erodes trust.

Recommend: remove the language enforcement from the default patch template. Document it in llm-anti-patterns.md as a pattern users can adopt, not a default injection. Or gate it with a flag like --enforce-language zh-CN.

⚠️ 2. integration.sh exists but is silently excluded from run_all.sh

tests/integration.sh is present in the branch but tests/run_all.sh only calls the other 3 suites. The CI step runs run_all.sh, so integration.sh never executes in CI.

Either include it in run_all.sh, or add a comment explaining the exclusion.

⚠️ 3. unit_codegen.sh only syntax-checks gen-presets.py, not execution

python3 -m py_compile "${REPO_ROOT}/scripts/gen-presets.py"

This only validates the AST. Runtime bugs (wrong file path, regex mismatch on YAML structure, marker not found in README) silently pass CI. Since gen-presets.py is a published workflow step, the test should actually run the script in a temp dir and assert output files exist and contain expected role blocks.

Minor (non-blocking)

• .worktree-task.md in the worktree root (Scope list) still references docs/gotchas.md which was moved to docs/gotchas/README.md in this PR.
• gen-presets.py uses hardcoded CWD-relative paths. Running from a subdirectory silently does nothing. A guard like os.chdir(repo_root) or a --root flag would make it more robust.
• Deleting docs/backlogs/priorities.md and folding into backlogs/README.md is clean — no dangling refs found.