fix(compile): pre-create the XLA cache dirs group-writable — first-creator ownership locked other users out

[claude-spd1]

Description

The shared $PARAM_DECOMP_OUT_DIR/xla_compilation_cache is multi-user, but jax >= 0.10 auto-enables the XLA per-fusion autotune side cache under it, and XLA creates xla_gpu_per_fusion_autotune_cache_dir/ and its tmp/ itself with hardcoded 0755 (umask-independent — that's why the files inside were 664 but the dirs weren't group-writable). Whichever user's run touched the cache first locked every other user out: their runs died with PERMISSION_DENIED writing tmp_per_fusion_cache_*.textproto at the first train step.

New param_decomp/compile_cache.py::ensure_group_writable_cache_dirs pre-creates the cache tree group-writable before jax initializes the cache; both jax entrypoints call it (LM trainer on rank 0, pretrainer in _maybe_enable_compilation_cache). Dirs that already have group write are untouched; a broken dir owned by someone else now fails fast at startup instead of one ~25-min compile later.

Note for #: bridge/task-jax-worker-persistent-cache introduces a param_decomp/compile_cache.py of its own (the one-policy-definition refactor). Same module name is deliberate — on merge, combine both functions in that module and call ensure_group_writable_cache_dirs from its enable_persistent_compilation_cache.

Related Issue

Bridge task xla-autotune-cache-group-unwritable (run p-9e6a4af3 / job 1052115 was the observed casualty). The live shared dir was already healed by hand (swapped for a 2775 copy, 670 autotune entries preserved); this PR is the structural fix so first-creation never matters again.

Motivation and Context

Any non-first user's pd-lm run was dead on arrival at the first train step.

How Has This Been Tested?

• New param_decomp/tests/test_compile_cache.py: creates the tree group-writable, repairs a 0755 tree, leaves already-writable dirs untouched (3 passed).
• make check clean (ruff + basedpyright).
• Imported both entrypoints and ran the helper against the real healed production cache dir — verified it's a byte-for-byte no-op there.

Does this PR introduce a breaking change?

No.

🤖 Generated with Claude Code

Crew-Address: task/xla-autotune-cache-group-unwritable