cleanup: trim stale comments + attic README

Dropped dead job-ID narrative (job 60/64) on rollout_ablate_frac, the
'vanilla step 17' dead-run ref in eval.py, the 'old signed sum' dead-code ref in
proj.py, and the conversational 'current experiment line' lead. Removed the dead
probe-traj justfile recipe. Kept all TODO/FIXME and the 'why' memory-tuning
comments. Smoke green (cout->0).

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

justfile

@@ -298,10 +298,6 @@ queue-projected preset="full" vhack="out/vhack/v_hack_full.safetensors":
 probe-base-pool steps="20":
     uv run python -m projected_grpo.probe_distill --base-only --steps={{ steps }} --n-problems={{ steps }}
 
-# Trajectory comparator for the warmup-gen runs (vanilla vs projected).
-probe-traj:
-    uv run python -m projected_grpo.probe_traj
-
 # Print the results table prototype.
 table-proto:
     @cat docs/table_proto.md

scripts/attic/README.md

@@ -0,0 +1,9 @@
+# attic
+
+Parked, not deleted. Superseded exploration kept only because the persona-pair
+methodology may get cited in the paper appendix.
+
+- `make_pairsets.py`, `make_dataset_pairsets.py` — persona contrastive-pair
+  authoring (tasks #123-126, done). The live extraction path is
+  `pairs.PAIRS` (hand pairs) or `pairs_from_pool` (pool-derived). No justfile
+  recipe builds these anymore.

src/projected_grpo/eval.py

@@ -25,8 +25,8 @@ def ref_logprobs_via_zero_delta(
 
     Uses `logits_to_keep=L_c+1` so HF's lm_head only runs on completion-side
     hidden states; prompt-side logits never materialize. Saves
-    ~plen/(plen+L_c) memory at the lm_head call (~33% at plen=500, L_c=1024).
-    That was the OOM site at vanilla step 17 (long prompt -> 4 GiB lm_head spike).
+    ~plen/(plen+L_c) memory at the lm_head call (~33% at plen=500, L_c=1024) --
+    a long prompt can spike the full-logits lm_head ~4 GiB and OOM without this.
     """
     saved = {n: info["delta_S"].data.clone() for n, info in wrappers.items()}
     try:

src/projected_grpo/proj.py

@@ -37,9 +37,9 @@ def _hackward_cos(c: Float[torch.Tensor, "k"], gn: torch.Tensor) -> float:
 
     relu BEFORE aggregating is the point: the one_sided projection removes only
     relu(c) (the hack-ward axes), and with V orthonormal ||removed|| = ||relu(c)||,
-    so this reads directly as "fraction of the grad the projection strips". The
-    old signed sum(c)/||g|| let +/- axes cancel, so it read ~0 even while a large
-    hack-ward magnitude was being routed -- a misleading gauge of routing activity.
+    so this reads directly as "fraction of the grad the projection strips" (a signed
+    sum would let +/- axes cancel and read ~0 even while routing a large hack-ward
+    magnitude).
 
     After a one_sided erase, V @ g_proj = min(c, 0) (positive axes zeroed), so
     relu of it is 0 -> cos_post == 0 exactly. That clean SHOULD (cos_post -> 0) is

src/projected_grpo/train.py

@@ -146,15 +146,13 @@ class Config:
     # saturation -- if on-policy sampling collapses onto hacking, every rollout routes
     # to the quarantine and the deployed δS never sees a solve gradient. Grading these
     # ablated rollouts is also a free per-step deploy proxy (hk_abl / slv_abl).
-    # DEFAULT 0 (off). job 64 (frac=0.5) leaked: deploy hack climbed 0.00->0.125 over
-    # training while the no-floor job 60 held deploy hack=0.000. Mechanism: sampling
-    # from the DEPLOYED model makes its reward an optimization target whose optimum is
-    # hacking, so whenever the (stale) routing gate misses, those deploy-generated
-    # hacks update the kept knob -- the floor optimizes the very policy the quarantine
-    # is meant to keep clean. Neither gradient-routing paper samples from the ablated
-    # model during training (they sample knob-on, ablate at eval, let absorption
-    # localize). So default off; re-enable only to TEST the floor+fresh-gate hypothesis
-    # (does refresh-every-1 keep the gate sharp enough that the floor is safe?).
+    # DEFAULT 0 (off): sampling from the DEPLOYED model makes its reward an
+    # optimization target whose optimum is hacking, so whenever the (stale) routing
+    # gate misses, those deploy-generated hacks update the kept knob -- the floor
+    # optimizes the very policy the quarantine is meant to keep clean. Neither
+    # gradient-routing paper samples from the ablated model during training (they
+    # sample knob-on, ablate at eval, let absorption localize). Re-enable only to test
+    # whether a fresh gate (refresh-every-1) keeps the floor safe.
     rollout_ablate_frac: float = 0.0
     # Which grader flaw + factual hint this run trains on (a "hack class"). Sets
     # the prompt hint (HINT_REPLACE_TO) and how `passed` is graded in rewards.py.
@@ -260,8 +258,8 @@ class FastConfig(Config):
     at pp=4 x 20 steps)."""
     model: str = "Qwen/Qwen3-4B"
     steps: int = 60                   # 60 lets the lp_s-lp_t gap open at convergence
-    # current experiment line: 4-mode substrate pool + prog_wide persona pairs are the
-    # default so real runs need only --intervention (+ optional seed/refresh/mask).
+    # 4-mode substrate pool + prog_wide persona pairs are the default, so real runs
+    # need only --intervention (+ optional seed/refresh/mask).
     teacher_pool_dir: Path | None = Path("out/pools/substrate")
     vhack_pairs_path: Path | None = Path("out/pairsets/prog_wide.json")
     group: int = 8                    # G=8 so the locked-in mix_ratio=0.125 gives 1 teacher / 7 student