arm_a_scaling emits nothing on a c=1-only receipt and the gate still returns VERDICT PASS #2830

arm_a_scaling in scripts/perf_gate.sh emits zero lines on a receipt whose only band is c=1, and the gate still returns VERDICT PASS.

The mechanism

for c in sorted(bands):
    if c == 1:
        continue
    ...

c=1 is the denominator of scaling_efficiency(c) = (agg(c)/agg(1))/c, so skipping it is correct. But with only c=1 present the loop body never executes, the arm prints nothing, and nothing downstream notices its silence.

Observed in a W1 pre-flight on lambda, --phase merge, real receipt:

PASS   ArmC integrity
SKIP   ArmC-sig  (release-phase rule)
REPORT ArmB c=1 UNMEASURED (Arm A still gates this cell)
REPORT ArmD instrumentation absent: kv.*
SKIP   ArmE (W1)
VERDICT PASS host=lambda phase=merge workload=W1     rc=0

Arm A — the primary release-blocking arm, and the one the matrix calls "comparator-free, so it is the primary gate" — contributed nothing to that PASS.

The opposite polarity IS armed, which is what makes this a gap rather than a design

A c=16-only receipt gives FAIL ArmA band c=1 absent. So:

Arm A fails when its denominator is missing and stands down when everything but the denominator is missing. That asymmetry is the defect.

Why it has not bitten yet

cell_completeness catches a short band set — but it runs only at --phase release. At --phase merge, which is what a PR runs, nothing does. So a merge-phase PASS on a one-band receipt is not evidence Arm A passed, and there is currently no way to tell the two apart from the gate's output.

Suggested shape

Arm A should be unable to return without saying something. Either:

1. it emits an explicit SKIP ArmA (no band > 1 in this receipt) line, so silence becomes impossible; or
2. it FAILs a receipt that declares bands containing only c=1, on the grounds that such a receipt cannot exercise the arm it is nominally gated by.

(1) is the smaller change and preserves the merge-phase REPORT semantics. (2) is stronger but may collide with legitimate single-band diagnostic runs.

Whichever ships needs a fixture in --selftest for the c=1-only case specifically. The existing selftest is 43 passed / 0 broken and covers Arm A's REPORT path via expiry_w1_fixed_date_still_reports — but no row exercises the empty-loop case, which is exactly why a silent arm survived a green selftest.

Context

This is the seventh instance today of one defect class: a check that cannot report its own failure. The others were an EPIPE-inverted grep (105 sites, #2804), a chown swallowing its errors, a job timeout naming no step (#2811), five make targets advertising CI that no workflow invokes (#2808), a sed draining the stream its sibling needed, and a self-test satisfied by its own vacuity floor (#2829).

Refs APR-PERF-GATE-001 (#2706)