test(integration): make the utf-bom defaultFileEncoding test deterministic - #13196

Head 319c941 is fully green now — every check has completed with 0 failures, nothing running. The earlier "Downgraded from Approve to Comment: CI still running" looks like it raced a check that was still pending at that moment. Could a maintainer re-run @qwen-code /triage so the approval can be re-posted? (Review ledger for the downgrade shows "findings":[] — no outstanding code issues.)

Verdict: merge-ready — local maintainer verification, 26 verification cells executed, 0 unexpected failures. Verified head 319c941f3d05b36990ef9d411515948fbf76fde8 against base 47463b79a7dcf1559d03fd06611e39c7a2dec7d5 (= head parent; single-commit PR).

Central claim

The defaultFileEncoding case — the only real-model round-trip left in utf-bom-encoding.test.ts, and the test whose 3 × 300s timeouts blocked the v0.24.8-preview.0 release (#13066) — now runs deterministically against a fake OpenAI server in seconds, while the real write_file tool still writes the file, so the byte-level BOM assertions are unchanged.

A/B load-bearing proof

Two worktrees that differ only in the PR commit, each independently installed (--frozen-lockfile) and built (npm run build && npm run bundle). Same targeted command on both, in an identical credential-free environment (empty HOME, no OPENAI_*/QWEN_* variables):

The base arm's red is the bug's own signature in local form: the old test cannot pass at all without a live authenticated model, and with one it costs a real round trip per attempt — which is exactly what exceeded the 300s vitest timeout three times in a row on the loaded Docker release runner (quoted from release run 36602108252):

× BOM with defaultFileEncoding configuration > should create new file with BOM
  when defaultFileEncoding is utf-8-bom  900131ms (retry x2)
→ Test timed out in 300000ms.

Control cleanliness: readlink of node_modules/@qwen-code/qwen-code-core resolves into each tree's own packages/core on both sides — no cross-tree leak. (Head was additionally run with a scrubbed HOME — no ~/.qwen/settings.json — and still passes, since runForcedToolCallScenario passes explicit CLI flags that outrank user settings.)

Full-file parity (collateral sweep)

The whole file (8 tests) run on both arms under the same credential-free environment:

The red→green delta is exactly the converted test. The 2 remaining reds are the pre-existing real-model preservation cases (should preserve UTF-8 BOM when editing... / ...when overwriting...), untouched by this PR, failing identically on both arms with the same no-auth signature (the author explicitly defers them to a follow-up; they are Windows-skipped by design and were not the release blocker).

Mutation matrix (vacuity check)

Two mutants applied to the converted test at the head tree, one at a time; expected outcome for both: killed (red).

M1 is the load-bearing one: it proves the converted test still exercises the full chain settings.defaultFileEncoding → real write_file tool → EF BB BF bytes on disk, i.e. the scripted fake write_file call does not make the test vacuous. After each mutant the file was restored, git status verified clean, and the pristine rerun is green (that rerun is image 01).

Targeted gates (head 319c941f3d)

ESLint liveness was proven before citing it: a planted const UNUSED_PROBE = 42; was reported (@typescript-eslint/no-unused-vars), then removed and the file re-verified clean.

Findings

None. One observation, not a code issue: the PR also adds afterEach(() => vi.unstubAllEnvs()) to the second describe block, matching the first block — this matters because runForcedToolCallScenario stubs env vars, and the new test is that block's only vi.stubEnv consumer; the addition is correct and consistent.

Not covered

• Docker sandbox lane — this round ran on macOS (darwin/arm64) with QWEN_SANDBOX=false; the release-blocking lane is Linux/Docker. The mechanism under test (model round trip vs scripted fake call) is platform-independent, and the PR's own CI (incl. Integration Tests (no-AK, No Sandbox)) is green.
• The two pre-existing real-model preservation cases — untouched by this PR; verified only for cross-arm parity (identical no-auth failure on both arms).
• Windows — the whole file is Windows-skipped by pre-existing design.

Methodology

macOS (darwin/arm64), Node v22.23.1. Head tree: git worktree at 319c941f3d; base tree at 47463b79a7 (head's parent). Each tree: corepack pnpm install --frozen-lockfile → npm run build → npm run bundle. Test env: env -i HOME=<empty tmp> PATH=$PATH QWEN_SANDBOX=false CI=true (credential-free; CI=true only extends waitForToolCall polling, never reached on red arms since the CLI exits immediately). Commands: npx vitest run --root ./integration-tests cli/utf-bom-encoding.test.ts [-t ...]. Mutants: edit at head tree, rerun, git checkout -- restore, git status verify. Counts: 2 targeted head runs + 2 targeted base runs (expected-red) + 8 full-file head cells + 8 full-file base cells + 2 mutants + 4 gates (eslint probe, eslint clean, prettier, tsc) + 2 workspace-link realpath checks = 26 pass, 0 unexpected fail.

— local verification by @wenshao