feat(benchmarks): full-roster cold re-run, and enforce the cold protocol in the driver

[Xore]

The matrix mixes two regimes, and #1947 rule 2 says two vintages must never share
a table:

	models	regime
phases 1+2	89	contended (hp-llm-worker live), N=2 + escalation
phase 4	5	cold, workers stopped, N=3, every cell ±0

The #3023 cold probe showed contention did not move scores — a fully resident
control reproduced [63,63] exactly, and an escalated spilling cell resolved to
the majority the protocol had already chosen. So this is not a correction of wrong
numbers.

It makes the regime uniform, which matters because the top of the field is
separated by one point across sixteen models. A promotion decided on that margin
cannot rest on a mixed protocol.

N=2, not N=3

Part 4 found "22.1 of the 37.4 model-wall hours re-derived identical bytes" and
concluded repeats belong on the axes that are not fixed — seed, quant, prompt.
#3036 now escalates 2 → 3 → 5 automatically wherever runs actually disagree, so
genuine variance is still caught without paying a third run on every deterministic
cell. About a third cheaper for the same information.

sweep_extra.sh gains two things it should always have had

STOP_WORKERS stops hp-llm-worker and ghidra-revdeck-1 for the run and
restores them via a trap on EXIT/INT/TERM. #3023's finding was not that
contention broke the scores — it was that the protocol depended on an operator
remembering. This moves it into the driver. The trap covers INT/TERM because
this script is routinely killed between models.

KEEP_WEIGHTS_ABOVE_GB gates the post-model ollama rm on free space. That
removal was correct at 92% full and is actively harmful with terabytes free: it
destroyed all ten of the requant plan's source models, and it is precisely why a
cold re-run of 89 models has to re-download 80 of them.

Results go to 1947cold/, not over 1947full/

The contended numbers are not garbage. Contended-vs-cold across 89 models is the
largest evidence anyone will ever have on whether the regime matters — discarding
it to save disk would repeat the mistake that lost the weights in the first place.

The chaining avoids a GPU race

chain_cold.sh waits on a positive condition — every tag in
models_requant.txt has both tier files — rather than "the GPU is idle".
chain_phase3.sh is already armed on the same card; two idle-checks would fire in
the same poll window and double-book it. coldrun.sh keeps its own running-sweep
guard as a second line, because the failure mode here is a wasted day of GPU
rather than an error message.

Verified live

• coldrun.sh aborts while phase 2 is running
• chain_cold.sh aborts at 0/7 ladder tags
• the combined roster de-duplicates to 96 unique entries across the phase 1,
  phase 2 and ladder lists (case-insensitively, since Ollama resolves names that
  way and the rosters spell some quant tags differently — Two roster models are unpullable: Ollama rejects a model-name segment longer than 80 characters #2738)
• bash -n clean on all three

Refs #1947, #3023, #2245, #3031, #3036, #2641

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None