chore: bump interp-engine to 1.7.3 in inference, graph and nla

AGENTS.md

@@ -107,7 +107,7 @@ live here at `interp-engine/`; it is now its own repository,
 with its own tests, docs, lint gates and release process. Its comparison harness against
 TransformerLens, nnsight, vLLM and SGLang went with it.
 
-Each of the three apps pins an **exact version** (`interp-engine[...]==1.7.0`) in its
+Each of the three apps pins an **exact version** (`interp-engine[...]==1.7.3`) in its
 `pyproject.toml`. Exact rather than floored because the engine is first-party and moves in lockstep
 with these services: a relock for some unrelated dependency must not silently swap the code that
 serves every activation.

apps/graph/pyproject.toml

@@ -12,7 +12,7 @@ dependencies = [
   "pydantic>=2.11.4",
   "torch>=2.11.0",
   "transformers>=4.57.1",
-  "interp-engine==1.7.0",
+  "interp-engine==1.7.3",
   "python-dotenv>=1.1.0,<2",
   "hf-transfer>=0.1.9,<0.2",
   "psutil>=7.0,<8",

apps/inference/pyproject.toml

@@ -49,10 +49,10 @@ dependencies = [
   # 1.6.0 is additive on the surface this service calls: speed and memory work (half the
   # static-buffer VRAM on vllm-static), per-model point fixes for Gemma 4, and a new
   # `interp_engine.memory` sizer nothing here imports yet.
-  # 1.7.0 adds GLM-4 to the fused gate/up table, which is what lets mlp_pre and
-  # mlp_pre_linear serve on that family. Nothing else in the package changed, and no
-  # model served here is affected.
-  "interp-engine[vllm,quant]==1.7.0",
+  # 1.7.3 tags a system turn the chat template injects on its own (Llama's date
+  # preamble) as `role=system` with no message index, so the lens and NLA chats can
+  # show it as its own bubble. 1.7.1-1.7.2 are capture-side fixes.
+  "interp-engine[vllm,quant]==1.7.3",
 ]
 
 [dependency-groups]

apps/nla/pyproject.toml

@@ -23,7 +23,7 @@ dependencies = [
     # the vendored sglang stack. The vLLM version pin lives in the engine.
     # Pinned exactly, not floored: the engine is first-party and moves in lockstep with this
     # service, so a relock for an unrelated dependency must not swap it. Bump deliberately.
-    "interp-engine[vllm]==1.7.0",
+    "interp-engine[vllm]==1.7.3",
     # Weight-only quantization for the source model and reconstructor backbone
     # (--fp8-source / --fp8-reconstructor / --int4-reconstructor). Used to arrive
     # transitively through `sglang[all]` (which pinned torchao==0.9.0); with sglang