node-v0.8.66

[2026-05-17] chore(release): 0.8.66 lockstep (2eb074e)

Bundles three sub-releases since 0.8.65:

• fix(rewards-vault) e18b15bc — claim_rewards instruction was
  sending 7 accounts; contract expects 8 with pause_state at slot 1.
  Anchor was mapping reward_account into pause_state slot →
  AccountDiscriminatorMismatch (error 3002 / 0xbba). Operators
  could not claim work-order rewards from the CLI or node-ui. Extracted
  builder to pure helper + 12 regression tests.
• fix(heartbeat) 791f26de — LORA_TRAINING_MEM_FLOOR_MB 4096→8192
  and DILOCO_TRAINING_MEM_FLOOR_MB 6144→14336. mps fp16 path on
  Apple Silicon holds the full base model in unified RAM
  (Qwen2.5-7B ≈ 14 GB). Old floors let 16 GB Macs auto-classify as
  DiLoCo-capable then OOM-killed mid-weight-load. Architectural
  invariant flipped from GPU == LORA < DILOCO to
  GPU < LORA < DILOCO. GPU_TRAINING_MEM_FLOOR_MB (4096) kept low
  so under-resourced nodes still micro-train.
• test 8225f478 — realigned 3 suites that drifted from 0.8.55+
  implementation changes: install.spec.ts (brew + curl Vina flow),
  hardware.test.ts (homedir mock for transitive python-venv import),
  sprint-e-part1.spec.ts (resolvePython spawn arg). 131/132 suites
  green (1 skipped intentional). Test-side fixes only — no business
  logic touched.

Operator impact: Mac M-series nodes with <14 GB free RAM will stop
advertising diloco_training; <8 GB free stop lora_training on the
next heartbeat (~60 s after deploy). Coord skip-routes accordingly.
RunPod pods (32-128 GB RAM) unaffected. claim-rewards flow on
node-ui and CLI now succeeds.