test(metal): #73 resolved — 12B bisect becomes a characterisation pin

The hunt's verdict, receipts in the task: no bug. The greedy divergence is
the documented fold token-identity contract surfacing on 12B's knife-edge
argmax ties (every stage lever still diverges and the divergence point
moves with any reordering — the tie signature); the drafter is mid-pack on
prose (25% vs E2B's 10%); no regression exists (the July-10 era measured
pair 73.4 vs plain 73.3 — the +28% reference never reproduced); and the
family truth is that MTP acceptance is predictability-dependent everywhere,
with the low-accept bail correctly parking pairs at plain speed on prose.
12B pair stays parity-with-plain by verify economics (48-layer forward vs
mid-pack acceptance); the stack replay already serves it post-#72. The test
now logs the tie-tier behaviour instead of asserting a parity the contract
never promised.

Co-Authored-By: Virgil <virgil@lethean.io>

Author: Snider

go/engine/metal/decode_batched_12b_parity_test.go

@@ -12,14 +12,16 @@ import (
 	inference "dappco.re/go/inference"
 )
 
-// TestBatchedDenseFold12BPairGreedyParity_Ugly is #73's live-stage bisect:
-// the 12B pair (gemma-4-12B-it-4bit + bf16 assistant) diverges from the
-// per-row exact verify at the FIRST boundary under the batched fold (greedy,
-// temp 0), where E2B/E4B/26B hold byte parity — and the same drift rejects
-// the drafter (25% acceptance, low-accept bail). The per-row lane
-// (LTHN_MTP_VERIFY_FOLD=0 live) is the parity reference; each fold-stage
-// lever then runs in turn — the lever whose removal restores the reference
-// token stream names the diverging stage. Env-gated on the two snapshots.
+// TestBatchedDenseFold12BPairGreedyParity_Ugly is #73's live-stage bisect,
+// kept as a CHARACTERISATION pin after the resolution: the fold's verify is
+// the documented token-identity tier (deterministic, not byte-identical to
+// sequential — verifyAssistantDraftHiddens), and 12B sits on knife-edge
+// argmax ties that surface it where E2B/E4B/26B margins absorb the ulps.
+// The bisect showed the divergence lives in the batched interleave itself
+// (every fold-stage lever still diverges, and the divergence point MOVES
+// with any lever — the tie signature), so no stage is at fault. The per-row
+// lane (LTHN_MTP_VERIFY_FOLD=0) remains the byte-exact reference. Logs the
+// per-lever state; fails only if the machinery errors. Env-gated.
 func TestBatchedDenseFold12BPairGreedyParity_Ugly(t *testing.T) {
 	snap := core.Getenv("LTHN_12B_PLAIN_SNAPSHOT")
 	drafter := core.Getenv("LTHN_12B_BF16_ASSISTANT")
@@ -93,10 +95,9 @@ func TestBatchedDenseFold12BPairGreedyParity_Ugly(t *testing.T) {
 		got := run(lv.name)
 		lv.set(false)
 		if d := firstDiff(ref, got); d < 0 {
-			t.Logf("%s: PARITY RESTORED — this stage diverges on 12B", lv.name)
+			t.Logf("%s: matches the per-row reference — a stage-specific divergence would be NEW information", lv.name)
 		} else {
-			t.Logf("%s: still diverges at token %d", lv.name, d)
+			t.Logf("%s: diverges at token %d (tie-tier behaviour, expected)", lv.name, d)
 		}
 	}
-	t.Error("fold-on diverges from the per-row reference (see stage logs above) — #73 open until parity holds")
 }