hyperloom custom workload: the result can never say whether the graph stayed on HRX -- placement is null in all 147 artifacts #1008

Class: the artifact does not describe the measurement (same as #892, #930, #932, #947, #954, #977, #993 — all fixed),
but this one is a field that already exists and has never once held a value.

bench_summary.placement is null in 147 of 147 artifacts on this box (every result carrying a bench_summary), and
0 of the 71 serve.log files beneath ~/hyperloom-ws contain the line the parser is written for:

$ python3 - <<'PY'   # counts over ~/hyperloom-ws, 2026-11-04 13:2x
files scanned 7102  with bench_summary 147  placement null 147
PY
$ grep -rl "off HRX" ~/hyperloom-ws --include=serve.log | wc -l
0

placement_summary()/PLACEMENT_RE work — fed the engine's own documented line (docs/serve.md,
tests/hrx_placement_test.cpp:39):

placement_observer: placement: 512 nodes, 34 off HRX (MUL_MAT x 30, GET_ROWS x 4), 7 graph splits
-> {'graphs_reported': 1, 'nodes': 512, 'off_hrx': 34, 'worst_graph_off_hrx': 34}

The measurement simply never asks for the report. GGML_HRX_REPORT=2 (set by serve_measure) only makes ggml-hrx log
the counts, and llama-server maps the library's INFO lines to its trace level, so they reach serve's output only below
-lv 4. serve is the one that passes -lv 4, for --placement-report (and --strict-device)
(app/serve.cpp:1059, app/hrx_strict.h:60), and the workload passes neither. The comment that used to sit on that env
line asserted the opposite — "1bit serve exposes no -lv passthrough" — which is how the field stayed empty for the
whole series. (That sentence also went into #993's body; it is wrong and it is corrected there.)

Consequence: a result cannot say whether the graph it measured stayed on HRX. On this stack a node off HRX is a bug to
report (docs/hw-gate.md), and the one place the workload would say it is the field that is always null.

Wanted: the serve leg asks for the placement report, so every artifact records
{"graphs_reported", "nodes", "off_hrx", "worst_graph_off_hrx"}, and "not captured" (null) stops being the only value
the field ever has.

Note for comparability: --placement-report lifts the backend's verbosity, so absolute numbers after it are not directly
comparable with the numbers of record (measured at serve's default verbosity); it applies to the baseline and every
candidate alike, so lever deltas stay valid.