Clock pinning is UNAVAILABLE inside an `rc` lease: `nvidia-smi -lgc` returns LGC_RC=4 as root, and every clock-pinned figure in the records was taken over the retired ssh+flock path

[localai-bot]

Measured 2026-08-19 on dgx:gpu0, in three separate rc run jobs, each running as root in the worker pod:

$ nvidia-smi -lgc 2190
The current user does not have permission to change clocks for GPU 0000000F:01:00.0.
LGC_RC=4

The container is root but lacks the capability the driver requires.

Why this outlives the campaign that found it

.agents/benchmarking.md §"The clock is part of the measurement" instructs "Pin the clocks before measuring, under the lock", and offers sudo nvidia-smi -lgc 2100 as the recipe. That recipe works on the host path. Every clock-pinned figure in this repository was taken over the host + ssh + flock path, and AGENTS.md now forbids ssh to a fleet device. So the migration to rc leases silently removed clock pinning, and no record said so.

Same class as #1265: a capability the records assume, which the current access path does not provide. It is visible debt rather than a surprise now, but it was a surprise on 2026-08-19.

What it costs, measured

Inside a lease the SM clock can only be SAMPLED. tools/bench/gpu_clock_state.py still works and is the only attribution such a figure carries. On a thermally throttling GB10 that is not enough for a PAIR:

• Nine timed windows across two arms on 2026-08-19 recorded within-run SM-clock spreads of 12.92% to 26.36% against the 5% ceiling .agents/benchmarking.md sets. SwThermalSlowdown was active in every window; one also carried HwSlowdown + HwThermal.
• gpu_clock_state compare returned PAIRING_VERDICT=DISCARD on all three c1 pairings, even though the cross-arm rule passed perfectly — same boot id, both arms at a 2489 MHz median, median offset 0.0%.
• The consequence is that the Qwen3.8-27B bf16 c1 cell has two clean, complete absolutes from both arms and no ratio, and the ratio is owed rather than measurable. That is GPU time spent on a pair that could not be divided.

What would settle it

One of:

1. A lease that grants the driver capability -lgc needs (an rc worker change), which restores the documented method.
2. A recorded, ratified alternative: a thermal settle-and-hold procedure that gets the within-run spread under 5% without pinning, demonstrated rather than assumed.
3. An explicit decision that lease-measured pairs use a different clock rule, argued in a spec, with gpu_clock_state's ceiling reconciled to it. Not a widening of the assertion to make a red go green.

Records updated in the meantime: .agents/environment.md records the refusal beside the other measured lease capabilities, .agents/benchmarking.md names the exception where it instructs the pin, and .agents/benchmark-record.md carries the measurement.

Found while recording the Qwen3.8-27B bf16 c1/c8 re-measure (#915, #979). Not fixed in flow: the fix is an infrastructure capability this row has no authority over, and options 2 and 3 are each a spec of their own.