fix(bench): a cp1252 console must not be able to destroy a finished measurement

[niko86]

✅ Based on main — mergeable now. Follow-up to #476, from its first real run.

What happened

Both legs measured correctly. The Windows leg then died in print:

UnicodeEncodeError: 'charmap' codec can't encode character '−' in position 42

The Windows console is cp1252, which has no U+2212 MINUS SIGN — and the results
table header carried one: vs v2 (− = faster). Three release builds and sixty
timed runs, lost to a typographic character in a column heading.

Two changes, because the character was the trigger and not the fault

Ordering — this is the real defect. emit printed before it wrote. Both
file writes name their encoding, so only stdout was ever at risk; but printing
first meant a console fault also cost bench.log and left the artifact step with
nothing to upload. Durable outputs now go first — file, then job summary, then
stdout — so a presentation-layer failure can no longer take the results with it.

The stream, not just the character. The header loses its U+2212 (a hyphen
says the same thing), and stdout is reconfigured to UTF-8 so the em dashes and
arrows already in the verdict prose can't fire the same fault from a different
line. Fixing only the one character would have left the class open.

isinstance rather than hasattr on the guard: the duck-typed check narrows to
object, which ty correctly refuses to call a method on.

Verified against the condition, not a description of it

PYTHONIOENCODING=cp1252 reproduces the original traceback on the old code, and
runs clean here with bench.log written. The "before" is the genuine CI
traceback rather than a simulated one.

The measurement was not lost

The per-arm figures were in the run log. Applying the harness's own rule by hand:

	Linux (self-hosted)	Windows (self-hosted VM)
v2	300.1 ms	413.9 ms
v3	310.1 ms (+3.3%)	406.8 ms (−1.7%)
system (control)	408.7 ms (+36.2%)	672.9 ms (+62.6%)
noise floor	11.4 ms (3.8%)	33.9 ms (8.2%)
verdict	indistinguishable	indistinguishable

The control separated hard on both, so the instrument is demonstrably resolving
allocator effects — and against that, v3 does not beat v2 on either platform.
The two deltas even point in opposite directions, which is what a null result
looks like.

So narrowing the pin to cfg(target_os = "macos") would buy Linux and Windows
nothing measurable, and #448's unquantified premise is now quantified. The
Linux verdict is the harness's own; the Windows one is mine, computed by hand —
which is what this PR exists to stop being necessary.

Re-dispatch after merge if you want the harness to state Windows for itself; the
answer is not in doubt, only its provenance.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!