fix(tune): floor the bench freq median to whole kHz — awk sci-notation tripped the #62 throttle guard

[VijitSingh97]

Summary

_median averages the middle pair for an even sample count, and awk's OFMT prints large non-integral results (kHz clocks ~4.6e6) in scientific notation (4.6275e+06). The consumer's bf=${bf%.*} floor mangled that into 4 → 0 MHz → the #62 throttle guard falsely rejected healthy candidates, so tune --bench silently kept the seed baseline on real hardware.

Fix per the issue: floor at the one producer whose values are large (the freq writer in _xmrig_bench), via awk '{printf "%d", $0+0}' — $0+0 parses the sci-notation back to 4627500 and %d prints whole kHz. _median itself is unchanged (its other consumers — hashrate/watts — never exceed 1e6). The now-stale consumer comment updated; its .-floor kept as a harmless belt.

Tests

• Unit-style: replays a deterministic even sample pair (4627000/4628001 kHz) through the real _xmrig_bench freq-write path and asserts the file contains 4627500 (no e+).
• Black-box regression: a healthy ~4.6 GHz candidate whose clock jitters 1 kHz mid-window (FIFO-fed, deterministic even sample count) must not be flagged as throttled and stays eligible for adoption.

Both verified red against the unfixed writer (4.6275e+06 / falsely throttled: true), green after. make test: 1336 passed, 0 failed; shellcheck + shfmt clean.

Closes #266

🤖 Generated with Claude Code