grinder: local deepseek-v4-flash-dspark is now the DEFAULT tier

markus request: make local DeepSeek the default grinder until requested otherwise.
Was floor=[free] qwen/qwen3-coder with a cloud escalation rung; now a single local
tier. Paid [$] moonshotai/kimi-k3 stays MANUAL (GRIND_TIERS override) and
deliberately out of any automatic ladder so a grind cannot spend money on its own.

Also drops the noether systemd override that pinned the old ladder.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

bin/bullpen-grinder

@@ -18,11 +18,11 @@ BWRAP = shutil.which("bwrap")   # #76: sandbox the (model-generated) test run
 
 REPO   = os.environ.get("GRIND_REPO", os.path.expanduser("~/src/bullpen"))
 PROXY  = "http://hossenfelder.fritz.box:8082/v1/chat/completions"
-TIERS  = os.environ.get("GRIND_TIERS", "qwen3.6-coding,deepseek-v4-flash-dspark").split(",")  # floor -> ceiling
+TIERS  = os.environ.get("GRIND_TIERS", "deepseek-v4-flash-dspark").split(",")  # floor -> ceiling
 # @py coordinator points the remote grind here (free cloud models via hossenfelder; boltzmann-local is slow).
 # NOT big-pickle: it's an agent-harness model that returns null content on a bare completion. qwen3-coder
 # (code-specialized floor) -> gpt-oss-120b (bigger ceiling), both verified to route + emit FILE: blocks.
-PY_TIERS = os.environ.get("PY_GRIND_TIERS", "[free] qwen/qwen3-coder,[free] openai/gpt-oss-120b")
+PY_TIERS = os.environ.get("PY_GRIND_TIERS", "deepseek-v4-flash-dspark")
 MAX_ITERS   = 10          # total model attempts across all tiers
 ESCALATE_AT = 4           # attempts on a tier without progress before escalating
 COLLECTION_PENALTY = 500  # a module that won't import scores worse than any assertion-fail count