fix: add fmt check, fix symlink bug, isolate tools workspaces

[ChenxingLi]

Summary

This PR completes the remaining work from #3426 (after #3441 and #3442) and adds a fmt check phase to test.sh. The changes fall into two categories: fixes for existing bugs that surface in worktree environments, and a new fmt check that closes a gap in CI coverage.

Symlink bug in check_integration_tests

Before running integration tests, test.sh replaces target with a symlink to build/ so tests find the compiled binary. The original rm -f target cannot remove a real directory — the error is silently ignored, and ln -s build target creates a symlink inside the surviving directory rather than replacing it. Integration tests then run against a stale or missing binary. Changed to rm -rf target.

Workspace isolation for tools/*

The main workspace declares exclude = ["tools"], which works correctly in a normal checkout — Cargo walks up, sees the exclusion, and treats each tools/* package as standalone.

The problem appears in nested worktrees. When a worktree lives under the main repo (e.g. /conflux-rust/.claude/worktrees/xxx/), Cargo finds the worktree's workspace root first, sees the package is excluded, and continues walking upward — eventually reaching the main repo's workspace. At that level the package's relative path becomes .claude/worktrees/.../tools/<pkg>, which no longer matches exclude = ["tools"]. Cargo concludes the package is inside a workspace that doesn't claim it, and errors out.

Adding [workspace] to consensus_bench/Cargo.toml and evm-spec-tester/Cargo.toml makes each package its own workspace root, stopping the upward walk before it escapes the worktree.

Fmt check in test.sh

The fmt check was previously assigned to Jenkins Master, which lacks the nightly toolchain — so it was effectively not running. Moving it into test.sh restores the CI gate on Worker where the toolchain is available, and benefits developers who use test.sh as a local pre-commit check.

The check runs before Phase 1 (build) with independent anchors (=== Fmt check === / === Fmt check passed ===) to avoid renumbering the existing four phases.

SKILL.md monitoring updates

test.sh's check functions use tee /dev/stderr to echo captured output. When the log is collected via > file 2>&1, fd 2 points to a regular file, and tee's open("/proc/self/fd/2") creates a new file description at position 0 — overwriting earlier log content including phase anchors. The fix: the launch command now pipes through tee (2>&1 | tee file), so fd 2 inside test.sh is a pipe. Pipes have no seek position, eliminating the overwrite.

The monitoring guide also picks up the new fmt phase: the detection regex recognizes fmt anchors, and a pre-flight fmt check is added so formatting issues surface before the long build. The Phase 2 workspace resolution warning is removed, as the [workspace] fix above eliminates the underlying risk.

Test plan

Tested in a nested worktree at /conflux-rust/.claude/worktrees/fix-remaining-3426/.

Workspace isolation: cargo_fmt.sh -- --check hit the workspace resolution error on both consensus_bench and evm-spec-tester before the fix, and passed after.

Fmt monitoring: Launched test.sh via pipe with an artificially introduced fmt failure. The monitoring regex correctly detected Fmt check as the current phase and observed process exit before Phase 1. After reverting the failure, confirmed the phase transition from Fmt check passed to Phase 1/4.

Not yet tested: full end-to-end test.sh (build + integration + pytest) requires a clean build on Worker.

---

This change is 

[peilun-conflux]

fmt already checked in .github/workflows/lint.yml. We should only run it once in CI?