benchmarks: sweep_extra.sh escalates once and has no path for "still no majority" at N=3

[Xore]

Split out of a #3023 comment (round 7, batch A review). The comment itself says this is
"a gap in the escalation protocol itself, independent of this issue", and no issue
existed for it — filing so it is not lost when #3023 closes.

Benchmark/GPU-parked: this is in the excluded set for round 7 and should not be
worked until GPU access frees up.

The gap

sweep_extra.sh escalates a cell from N=2 to N=3 when run1 != run2, then stops.
There is no handling for the case where three runs produce three distinct values, so
no majority exists and the cell has no defensible score.

That case is live. From #1947 phase 2, the escalated cells:

A  codellama_7b-instruct-q4_K_M          [57, 58, 58] -> 58  (2/3)
A  qwen3_8b                              [64, 65, 65] -> 65  (2/3)
B  gemma-4-31B-it-qat-q4_0-heretic       [62, 63, 63] -> 63  (2/3)
A  pki/nova-24b-cybersec                 [58, 59, 58] -> 58  (2/3)
A  Foundation-Sec-8B-Instruct-Q8_0       [61, 62, 61] -> 61  (2/3)
A  Foundation-Sec-1.1-8B-Instruct-Q8_0   [62, 64, 62] -> 62  (2/3)
B  ravenx-cyberagent-35b                 [64, 63, 64] -> 64  (2/3)
A  GLM-4.6-REAP-218B:i1-IQ1_S            [55, 54, 55] -> 55  (2/3)
B  GLM-4.6-REAP-218B:i1-IQ1_S            [55, 57, 54] -> NO MAJORITY

Eight of nine resolve 2-of-3. The ninth does not resolve at all, and it is the most
CPU-offloaded row in the roster (57 GB served, 65%/35% CPU/GPU) — pointing at spill
nondeterminism rather than slot contention. It must not be published as 55, 57 or
54.

What would close this

1. sweep_extra.sh detects "N=3 produced no majority" explicitly instead of silently
   taking one of the values.
2. On that branch it either escalates to N>=5 or records the cell as UNRESOLVED.
3. The matrix renderer carries UNRESOLVED through to publication rather than
   dropping the row or picking a value.

Related

• ops(llm): #2582 has recurred — hp-llm-worker unhealthy for ~11h, and the #1947 phase-2 sweep is running contended, not cold #3023 — where this was measured
• docs(benchmarks): #1805-c ghidra-slot matrix — cold-protocol re-measure, and promote gemma-4-26B-A4B #2641 — the cold-slot protocol the escalation sits inside
• One benchmark, every model, every aspect: rebuild the model matrix on current hardware with error bars #1947 — the sweep this blocks a clean publication of

[Xore]

Closing: PR #3037 (merged 2026-09-05) delivered the core asks — sweep_extra.sh now detects a no-majority cell at N=3, escalates to N=5, and mark_unresolved_if_no_majority() writes UNRESOLVED_tier_.status markers instead of picking a winner. The remaining half (the matrix renderer must carry UNRESOLVED into publication, originally #3036's scope, since closed) is posted to #3087, which owns the round-7 write-up and promotion records.