ci(benchmark): surface head-build failures instead of hiding make output

[Levalicious]

Follow-up to #245 (held back because pushing workflow files needs the workflow token scope — now granted). On head-build failure, print the tail of the build log instead of sending make output to /dev/null — that suppression is why the #244 red run was undebuggable. No behavior change on success.

🤖 Generated with Claude Code

[github-actions]

Per-op benchmark — p50 cycles (base vs head)

base runs: 5, head runs: 5 (median of mins) · freq-correction ×0.9972 · threshold head−base > max(40 cyc, 5%)

op	base	head*	Δ	Δ%
lookup	83.9	85	+1.1	+1.3%
read_entity	28.8	29	+0.2	+0.7%
neighbors_d1	398.2	408	+9.8	+2.5%
neighbors_d2	1565.4	1580	+14.6	+0.9%
find_path_d6	132223	142610	+10387.0	+7.9%	🔴 REGRESSION
search	800644	799374	-1270.0	-0.2%
entities_by_type	11024	11019	-5.0	-0.0%
random_walk_d5	833.8	834	+0.2	+0.0%
inc_walker_visit	12	12	+0.0	+0.0%
structural_sample	1845454	1850730	+5276.0	+0.3%
compute_merw_psi	583856	585634	+1778.0	+0.3%
create_entity	598	596	-2.0	-0.3%
create_relation	546	570	+24.0	+4.4%

* head frequency-corrected to base via the reference op.

Regressions detected.