CI: disable JAX GPU preallocation for concurrent notebook execution

Fixes the cold-cache kernel death diagnosed in the instrumented run
(30806359596). The three JAX lectures execute concurrently on the
g4dn runner's single T4; whichever kernel touches the GPU first
preallocates JAX's default 75% (11.3GiB of 15.3GiB). The diagnostics
showed jax_intro's kernel losing that race, clamping to the ~3.4GiB
remainder, and aborting ~40s later when a cell outgrew the pool --
an XLA fatal with no Python traceback, no OOM-killer or segfault
records in dmesg, and GPU usage pinned at 3427MiB until the death.

The jupyter-book-era green cold run won the same race by seconds --
the engine swap changed execution phasing, not semantics -- and the
JB1 production build on main executes sequentially, which is why
neither ever hit this.

XLA_PYTHON_CLIENT_PREALLOCATE=false switches JAX to on-demand
allocation so the concurrent kernels share the GPU. The memory
monitor and dmesg diagnostics stay in place; this run doubles as
the fix's validation.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: mmcky

.github/workflows/ci.yml

@@ -89,6 +89,16 @@ jobs:
       # errors, since ipynb export is a fork-only capability).
       - name: Build HTML
         working-directory: ./lectures
+        env:
+          # Three JAX lectures (jax_intro, autodiff, numpy_vs_numba_vs_jax)
+          # execute concurrently on the g4dn's single T4. By default the first
+          # JAX process to touch the GPU preallocates 75% of it (11.3GiB of
+          # 15.3GiB); the 2026-08-03 diagnostics showed jax_intro's kernel then
+          # clamped to the ~3.4GiB remainder and aborted with no Python
+          # traceback once a later cell outgrew that pool (dmesg had no
+          # OOM-killer or segfault records — an XLA fatal, not a host kill).
+          # On-demand allocation lets the concurrent kernels share the GPU.
+          XLA_PYTHON_CLIENT_PREALLOCATE: 'false'
         run: myst build --html --execute
 
       # A kernel death with no Python traceback is consistent with a SIGKILL