fix: improve monitor correctness and robustness

[staryxchen]

Summary

Four correctness/robustness fixes to nvlink_monitor, all identified by reading the existing source. No behavioral change to nvlink_bw_test.

Fixes

1. Async-signal-safe signal handler (monitor/nvlink_monitor.{h,cpp})

• g_running: volatile bool → volatile sig_atomic_t (the standard-mandated type for writes from a signal handler).
• signal_handler() no longer calls std::cout — output streams are not async-signal-safe and can deadlock if the signal interrupts the main thread mid-output. The handler now only flips the flag.
• The "Received stop signal, exiting..." notice is printed from main() after the monitoring loop observes the flag, covering both continuous and single modes.

2. Timestamp consistency in runContinuousMonitoring (monitor/nvlink_monitor.cpp)

The per-iteration timestamp was recorded before getNvLinkData(), while the pre-loop lastTime was recorded after its read. This made iteration 1's actualInterval exclude the read duration ε while iteration 2+ included ε + the previous iteration's calculate/print time, producing inconsistent bandwidth denominators.

Fixed by recording currentTime after getNvLinkData() returns, so both lastTime and currentTime mark "moment a counter read completed" and every actualInterval exactly equals the observation window between two reads (which is what NVML counters actually reflect).

3. NVML fallback counter unit caveat (monitor/nvlink_monitor.cpp)

The nvmlDeviceGetNvLinkUtilizationCounter fallback returns raw counters whose units depend on nvmlDeviceSetNvLinkUtilizationCounter config and are not guaranteed to be KiB like NVML_FI_DEV_NVLINK_THROUGHPUT_DATA_TX/RX. The code feeds them through the same KiB→GiB conversion regardless.

• Added a code comment documenting the caveat.
• Gated the per-link fallback debug print behind --verbose (it previously spammed stdout unconditionally every sample in continuous mode).
• The verbose warning now notes the bandwidth estimate may be inaccurate on this path.

4. Counter reset wording & handling (monitor/bandwidth_calc.cpp)

A negative delta on the 64-bit KiB throughput counters (~8 EiB wrap range) almost always indicates a driver counter reset, not arithmetic overflow. The old warning said "overflow", which is misleading.

• Renamed warning "overflow" → "counter reset" and included the delta value.
• Added a comment explaining no meaningful rate can be computed across a reset, so the sample is clamped to 0.
• Renamed test CounterOverflowClampedToZero → CounterResetClampedToZero.
• Added CounterResetVerboseWarningDoesNotCrash to exercise the verbose stderr branch.

Verification

• make check-format — clean
• make test — 59/59 pass (was 58; +1 new test)
• make monitor — builds with -Wall -Wextra, no warnings (NVML link verified)

Files changed

 monitor/bandwidth_calc.cpp   | 19 +++++++++++++-----
 monitor/nvlink_monitor.cpp   | 53 ++++++++++++++++++++++++++++++++++++++++++-----------
 monitor/nvlink_monitor.h     |  6 ++++--
 test/test_bandwidth_calc.cpp | 16 ++++++++++++++--
 4 files changed, 75 insertions(+), 20 deletions(-)