bench-log: the 9070 XT's four drops on 5 October and the kit wiped by the 0.3.10 install; the playbook notes the re-fetch rule

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/bench-log.md

@@ -1591,6 +1591,8 @@ Reading: the best point is the 400 W floor at 0.369 MH/W, 114.46 MH/s, a hash co
 
 **AMD clock and power control, and the 9070 XT sweep: OWED** (branch `opencl-rdna4-telemetry`; the project lead: "can anything be done to redeem AMD?"). The control is built and unit-tested, the measurement is not taken: the RX 9070 XT left PC 1's bus before the run. Relay probe #202 at 20:40 UTC listed only the gfx1036 and the RTX 5090 under Display; probe #203 at 20:45:34 UTC, after `pnputil /scan-devices`, the same, and the USB4 list had lost the "USB4 Router (2.0), Sonnet Technologies Breakaway Box 850T5" that was present at 17:18 UTC, so the box itself is off the link (the second eGPU fault of the day; the first, Code 43 at 17:18, needed a driver reinstall and a reboot). Nobody restarts PC 1 tonight; the project lead reseats the link in the morning. The app fell back to the gfx1036 (3.12 MH/s on `--device 1`). The PC 1 scheduler dropped the slot (job 4) at 20:46 UTC.
 
+Later the same night: the card came back by itself at about 21:09 UTC, was absent again at 21:22:59 UTC (probe #205), was listed and used by the era agent's job up to its 22:21 release, and was absent again at 22:24:09 UTC (probe #230, the scheduler's go for the sweep; Display listed only the gfx1036 and the 5090, no Sonnet router): four drops in one evening, so the slot was released without starting. The kit zip fetched at 21:23:59 UTC into `jobs\amd-kit-1\kit` was gone by 22:24: the 0.3.10 install at 21:49 clears the app's jobs folder, so the playbook's fetch job must be published again, under a new id, after any app update and before the run (`$KitJob` names it).
+
 What is ready to run the moment a discrete AMD card is on the bus: `relay/playbooks/amd-card-test.ps1` (one job; parameter `$Gfx`, `gfx1201` for the 9070 XT, `gfx1100` for the RX 7900 XTX the project lead has ordered), which switches that card off in the app, checks the `--list` fold (one row, no dup row), runs `--memprobe` at 4, 64, 96 and 1024 MiB, a 120 s hash run in the worker's own serve mode on `pack-a` (within 1% of the app's rate on the 9070 XT, above), the telemetry window, then the core-clock sweep (stock, 2800, 2400, 2000, 1600, 1200 MHz) and the power-limit sweep (100, 80, 65, 50% as ADLX offsets 0, -20, -35, -50, clamped to the card's range) at 90 s each with the memory clock recorded at every step, `--reset` to factory at the end with a `--tune` read-back and a 30 s window after it, and prints the table below filled, the best MH/W point with its hash cost, whether the memory clock held, and the MH per pound (`$PricePounds`). The rows this entry owes, with the expected stock values from the measurement above:
 
 | Setting | MH/s | Watts | MH/W | Temp C | Fan rpm | Memory MHz | Shader MHz |

relay/playbooks/amd-card-test.ps1

@@ -3,7 +3,7 @@
 # the kit zip (igneum-worker-opencl.exe, igneum-gpu-telemetry.exe, kernel_bound.cl, pack-a\) under jobs\<KitJob>\kit.
 # Parameters at the top (a job script takes no arguments; edit the three lines or sed them before publishing):
 #   $Gfx        the card's OpenCL name as the worker lists it (gfx1201 = RX 9070 XT, gfx1100 = RX 7900 XTX)
-#   $KitJob     the fetch job id whose kit folder holds the exes
+#   $KitJob     the fetch job id whose kit folder holds the exes (an app update clears the jobs folder: fetch again, new id, after one)
 #   $PricePounds the card's price for the MH per pound line (0 = not stated)
 # What it does, with the card under test switched OFF in the app (POST /api/cards, that key only; the other cards keep
 # mining) and switched back on at the end: the --list fold check (one row for the card, no dup row), --memprobe at