#2602 docs: INTEL.md: a card lost while reading a prompt with THP "always" and swap on, and the two keys that stopped it (#2157)

Beschreibung

The docs note #2157's reporter suggested.

**What it changes.** One bullet in `docs/INTEL.md` ("Things that matter on this GPU"), with #2157's machine and numbers only: 2x Arc Pro B70, xe, kernel 7.2.9-cachyos, 32 GB RAM, zram swap, `vm.swappiness=150`. `UR_RESULT_ERROR_DEVICE_LOST` while reading a prompt (`Timedout job` in dmesg); the reporter's trace shows the kernel moving the cards' system-memory buffers to swap mid-job and the card faulting them back past xe's 5 s job timeout; conversation-cache parks set it off in their tests (2 of 3 runs with a 12 GiB cache, 0 of 4 with the cache off or a 64 MiB budget); THP `madvise` 0 of 4 but lost again with `--pipeline-windows 2`; `NEOReadDebugKeys=1` + `EnableRecoverablePageFaults=0` in the engine's environment stopped it (0 of 4; decode 71.8 / 96.3 / 72.0 tok/s and the same greedy text on 5 prompts).

It also says where the keys work: an engine run directly gets both from the run config's `"env"`, but the Docker launcher setup uses (`sycl/serve/strata-sycl.sh`) forwards `NEOReadDebugKeys` and not `EnableRecoverablePageFaults` (its name list), and drops any value of `0`, so the second key does not arrive there yet. Changing the launcher is a separate (code) change; this PR does not make it.

**What it leaves alone.** The launcher, the conversation cache, and every other line. No new measurement: the numbers are the reporter's, marked as small samples.

**How it was checked.** Against the issue and its comments, and `strata-sycl.sh`'s forwarding loop. A sub-agent reviewer checked every number and claim; its corrections (the launcher gap, the `madvise` result per config, the hedged trigger, the decode numbers) are in.

🤖 Generated with [Claude Code](https://claude.com/claude-code)