[Bugfix][Store] Isolate fileread worker count assertions from runtime threads

[Aionw]

Description

Fix the flaky FilereadWorkerPoolTest.AcceptsTypedTrailingWhitespaceAndCaches introduced in #4166 and observed in the unrelated #4161 CI run:
https://github.com/kvcache-ai/Mooncake/actions/runs/35190325962/job/105191941422?pr=4161

The test counts every thread under /proc/self/task and requires the process-wide count to return to its original baseline after pool destruction. CI enables USE_ETCD and STORE_USE_ETCD, loading a Go shared library whose runtime can create unrelated background threads asynchronously. Joining all pool workers does not restore that process-wide baseline.

Add a narrow friend test peer to inspect the pool's own worker vector. Preserve both assertions: "2 " creates two workers, and changing the environment to "3" still creates two workers because the configuration is cached. Keep both construction/destruction scopes; remove process-wide thread polling and the now-unnecessary Linux-only guard. No production behavior or public API change.

Searched existing issues/PRs for the test name, fileread_worker_pool, and WaitForProcessThreadCount; no dedicated fix found. #4179 reorganizes store paths but does not address this assertion race.

Module

• Mooncake Store (mooncake-store)

Type of Change

• Bug fix

How Has This Been Tested?

Test commands:

CCACHE_DIR="$HOME/.cache/ccache/mooncake" \
CCACHE_BASEDIR="$(git rev-parse --show-toplevel)" \
cmake --build build --target fileread_worker_pool_test \
  fileread_worker_pool_config_test transfer_task_test -j 12

ctest --test-dir build \
  -R '^(fileread_worker_pool_config_test|fileread_worker_pool_test|transfer_task_test)$' \
  --output-on-failure
ctest --test-dir build -R '^fileread_worker_pool_test$' \
  --repeat until-fail:1000 --output-on-failure
# ETCD_WRAPPER_SO points to an existing local Go-built libetcd_wrapper.so.
LD_PRELOAD="$ETCD_WRAPPER_SO" ctest --test-dir build \
  -R '^fileread_worker_pool_test$' --repeat until-fail:1000 --output-on-failure

# Use clang-format 20 for the repository's version-pinned formatting hook.
uv tool run --from clang-format==20.1.8 sh -c \
  './scripts/code_format.sh --check --staged'
uv tool run --from clang-format==20.1.8 sh -c \
  'prek run --files mooncake-store/include/transfer_task.h mooncake-store/tests/fileread_worker_pool_test.cpp'

Test results:

• Unit tests pass: all three focused CTest targets.
• Integration tests pass (not run).
• Manual testing done: controlled before/after runtime-loading experiment.

Before the fix, the local CPU build (etcd disabled) passed 100/100 separate executions. Loading an existing libetcd_wrapper.so with LD_PRELOAD reproduced 13 failures in 100 separate executions: nine at line 76 (the exact CI assertion), four at line 70 (the equivalent first-pool assertion). After the fix, both plain and Go-runtime-preloaded CTest repetitions passed 1000/1000. This is a controlled reproduction, not a full recreation of the CI build; CUDA and etcd build options remain disabled locally. Existing build configuration uses ccache launchers and disables sccache.

All applicable touched-file hooks passed with prek (pre-commit unavailable) and clang-format 20.1.8. The initial formatting attempt with the host's clang-format 22 was rejected by the version-pinned script; rerunning with 20 passed.

Checklist

• I have performed a self-review of my own code (human review pending; draft PR).
• I have formatted my code using ./scripts/code_format.sh.
• I have run pre-commit on the files changed in this PR and all hooks pass (prek equivalent).
• I have updated the documentation (not applicable: test-only behavior).
• I have added tests to prove my changes are effective (updated existing regression test and repeated the reproducer).
• For changes >500 LOC: I have filed an RFC issue (not applicable).

AI Assistance Disclosure

• No AI tools were used.
• AI tools were used.

An OpenAI coding assistant investigated CI logs, reproduced the failure, implemented this focused fix, and ran the checks. This PR is intentionally a draft pending the human submitter's line-by-line review and confirmation that they understand and can defend the change end to end.

[fcczzz]

LGTM