proj: measure_only kwarg + train.py always-on cos_in diagnostic

Vanilla arm now reports cos_in per step too (cosine of accumulated
Dr.GRPO grad with v_hack), as long as v_hack file is on disk. The
projection action only mutates the gradient when arm=projected;
vanilla just measures.

This makes Phase 2 (pilot scale) directly inform Phase 3: vanilla
cos_in trajectory says whether v_hack is even aligned with the GRPO
direction, before we burn 65h on the full sweep.

src/projected_grpo/proj.py

@@ -25,6 +25,7 @@ def project_delta_S_grad(
     wrappers: dict,
     v_hack: dict[str, torch.Tensor],
     preserve_magnitude: bool,
+    measure_only: bool = False,
 ) -> dict[str, float]:
     """Per-module one-sided removal of v_hack-aligned component from delta_S.grad.
 
@@ -32,6 +33,9 @@ def project_delta_S_grad(
     If cos(g, v) > 0: g' = g - <g, v> v (remove projection onto v). Optionally
     rescale g' to ||g|| to preserve update magnitude. Else leave g untouched.
 
+    If `measure_only`: same cosine math, but the gradient is NOT mutated.
+    Used by vanilla arm to report cos_in trajectory as a diagnostic.
+
     Returns aggregate diagnostics: mean_cos_in, mean_cos_out, frac_fired.
     """
     cos_in_list, cos_out_list, n_fired = [], [], 0
@@ -53,7 +57,8 @@ def project_delta_S_grad(
                 g_proj = g_proj * (gn / gp_n)
             cos_out = (g_proj @ v) / g_proj.norm().clamp_min(1e-12)
             cos_out_list.append(cos_out.item())
-            info["delta_S"].grad = g_proj
+            if not measure_only:
+                info["delta_S"].grad = g_proj
             n_fired += 1
         else:
             cos_out_list.append(cos_in.item())

src/projected_grpo/train.py

@@ -317,13 +317,15 @@ def main(cfg: Config) -> int:
     delta_params = [info["delta_S"] for info in wrappers.values()]
     logger.info(f"trainable delta_S: {sum(p.numel() for p in delta_params):,}")
 
-    # v_hack only needed for projected arm. Vanilla H4 sanity runs do not
-    # require a precomputed v_hack and should not be blocked by missing one.
-    if cfg.arm == "projected":
+    # v_hack: loaded for both arms when the file is present, so vanilla also
+    # reports cos_in as a diagnostic (no projection applied). If not present
+    # and arm=vanilla, skip silently — H4 sanity runs without v_hack remain valid.
+    v_hack = None
+    if cfg.v_hack_path.exists():
         v_hack_cpu = load_v_hack(cfg.v_hack_path, model_name, wrappers)
         v_hack = {name: v.to(device) for name, v in v_hack_cpu.items()}
-    else:
-        v_hack = None
+    elif cfg.arm == "projected":
+        raise FileNotFoundError(f"projected arm requires v_hack at {cfg.v_hack_path}")
     opt = torch.optim.AdamW(
         delta_params, lr=cfg.lr, weight_decay=cfg.weight_decay,
         betas=(cfg.adam_beta1, cfg.adam_beta2),
@@ -535,9 +537,12 @@ def main(cfg: Config) -> int:
             agg_loss += loss.item()
             t_fb += time.perf_counter() - _tfb
 
-        # One projection on accumulated grads (projected arm only).
-        if cfg.arm == "projected":
-            diag = project_delta_S_grad(wrappers, v_hack, cfg.preserve_magnitude)
+        # Diagnostic cos_in for both arms; projection only mutates grad if arm=projected.
+        if v_hack is not None:
+            diag = project_delta_S_grad(
+                wrappers, v_hack, cfg.preserve_magnitude,
+                measure_only=(cfg.arm != "projected"),
+            )
         else:
             diag = {"mean_cos_in": float("nan"), "mean_cos_out": float("nan"), "frac_fired": float("nan")}