fix(langgraph): keep dev-runtime idle pickle writes off the host NVMe

[Exploitacious]

Stacked on #3 (base branch sync/upstream-2026-09-26). Fixes the idle disk abuse measured on VM 150.

Root cause (file:line, pinned versions)

langgraph dev always runs the in-memory runtime (it hardcodes :memory: and never reads DATABASE_URI). That runtime's daemon flush loop (langgraph_runtime_inmem 0.30.0 _persistence.py: _flush_interval = 10, _flush_loop) calls PersistentDict.sync() on every store every 10s unconditionally. sync() (langgraph_checkpoint 4.1.1 langgraph/checkpoint/memory/__init__.py) re-pickles the ENTIRE dict and atomically replaces the file, with no dirty flag anywhere in the class and no pruning. The blobs dict (.langgraph_checkpoint.3.pckl) is only pruned by an explicit delete_thread().

Measured on VM 150 (idle): .3.pckl reached 198 MB, rewritten every ~10s = 26.9 MiB/s, 32.3 TB in 18 days, 0.44%/day of NVMe wear.

Why the runtime's own disable switch does not work here (verified on this box)

LANGGRAPH_DISABLE_FILE_PERSISTENCE / disable_persistence cannot be set through langgraph dev in the pinned versions:

• langgraph-cli 0.4.29 validate_config_file() strips the unknown disable_persistence key (the validated config's keys do not contain it), so config_json.get("disable_persistence", False) is always False.
• run_server() then builds LANGGRAPH_DISABLE_FILE_PERSISTENCE=str(disable_persistence).lower() = false and applies it with patch_environment(), which refuses to overwrite from the loaded env. Observed directly: the worker subprocess env showed the flag false even when the container was launched with it true, and the .pckl files were still written.

There is also no knob for the 10s interval (_flush_interval is a hardcoded constant).

Fix

Back /app/.langgraph_api with a 1 GiB size-capped tmpfs in the compose langgraph service. The runtime still re-pickles every 10s, but into RAM, so the NVMe sees nothing. LANGGRAPH_DISABLE_FILE_PERSISTENCE=true is set as a forward-compatible no-op that takes over if upstream honors it. langgraph.json is left unchanged (an unknown key is stripped now and could be rejected by a stricter future validator). Design and rejected alternatives: docs/adr/0013.

Before/after (this box, one image build, idle ~60s each)

• Before (no tmpfs): .langgraph_ops.pckl (8.4 MB after one seeded run) rewritten every 10s; container block-IO write climbed to 277 MB.
• After (tmpfs mounted): same file still rewritten every 10s into RAM; tmpfs held ~8 MB; container block-IO write stayed 0 B. mount confirms tmpfs on /app/.langgraph_api.

Guard

packages/decepticon/tests/unit/ops/test_langgraph_persistence_config.py asserts the shipped compose mounts the size-capped tmpfs and sets the forward-compat flag. It fails on the old config (no tmpfs / no flag) and passes here.

Checks

• uv run pytest -n auto -q -m "not slow": 5179 passed, 44 skipped.

Trade-off / open item for maintainer

LangGraph run history no longer survives a container restart (the tmpfs clears). It already did not survive a recreate (/app/.langgraph_api was never volume-mounted) and the stack already ran :memory:; engagement evidence persists separately. Residual: the runtime still burns CPU/RAM re-pickling every 10s; only the disk cost is removed. The complete fix is upstream honoring the disable switch. Do not merge without maintainer review.