fix(qwen38): record at MNBT 1,024 — the step size the cache can actually pay for

The rerun after 96c5a06 refused every corpus document again ("document 196 is
8507 tokens and the KV pool holds 7436"), and the boot table shows why the
host-pool change could not have helped: vram_budget.cpp put the container in its
`no lender` branch — elastic 25.56 GiB, experts 25.29 GiB (their plan-fixed
share), kv cache 271.25 MiB — identical at 12,288 and at 49,152 MiB of host
pool. The only supplier left there is the activation arena, and the engine says
so at vram_budget.cpp:117: "lower --max-num-batched-tokens (the arena scales
with it)". The log prices both rungs directly (a 2,048-token step costs 1.08
GiB, a 1,024-token step 504.39 MiB), so the default moves to 1,024: the cache
grows to 866 MiB / 24,300 tokens and scores every one of the 197 documents
(longest 11,970). Calibration keeps its wide 65,536; it was never tied to MNBT.

This also retracts the mechanism 96c5a06 recorded in the header ("the pool buys
KV as well as residency"). Inheriting the env's 49,152 / 96 stays right as
single-source-of-truth, but on this payload the pool is not the cache's supplier.

scripts/run-qwen38-bf16.sh

@@ -39,37 +39,45 @@
 # speculation, fp8 KV, a 32,768-token window, the PLE n-gram table pinned to RAM,
 # and the production chat template rather than the one baked into the container —
 # every scored container is asked the same tokenised questions the reference
-# answered. The window is 32,768 because this corpus is long: measured on
-# Dankserver 2026-10-27, kld.jsonl is 197 records with a longest of 38,263
-# characters (longest prompt FIELD 36,855, median record 27,228), which is ~9,600
-# tokens at the pessimistic chars/4 and ~7,700 at the ratio this repo actually
-# measured on the K3 corpus (380,143 chars = 76,568 tokens, chars/4.97). A 16,384
-# window cleared that by 1.7x; 32,768 clears every record of both corpora by at
-# least 2.6x, and the kldref gate below re-checks it against the file on disk.
+# answered. The window is 32,768: kld.jsonl is 197 records with a longest of
+# 38,263 characters (longest prompt FIELD 36,855, median record 27,228). Its
+# token length — 10,554 for that longest record, counted by the engine itself —
+# is in the funding paragraph below; the kldref gate re-checks the file against
+# the window before it boots anything.
 #
 # NO --vram-kv-mib: it was REMOVED at 1.3.0 and naming it is a boot refusal, not
 # a no-op — verified on the pinned image 2026-10-27: "E --vram-kv-mib was
 # removed: the card is split between the weights and the cache from what the
-# model needs." The budget resolver derives the per-card split from
-# --host-pool-mib and --gpu-headroom-mib, and those two are deliberately NOT
-# production's 49,152 / 1,536 (frank-archon.env:77): the pool decides WHICH
-# experts stay resident, so it buys speed and never logits — the shape the gate
-# has to hold identical is the kv dtype, the slot count, the step size, the spec
-# depth and the window. It is also the only figure that fits: a 335 GiB bf16
-# container cannot be booted against a 49 GiB resident set on a 192 GiB machine,
-# and a prefill-only pass does not care that the disk tier paces at ~3 GiB/s.
+# model needs." The resolver derives the split from --host-pool-mib,
+# --gpu-headroom-mib and the activation arena; the rig inherits the first two from
+# frank-archon.env, the file that owns them. On THIS payload the host pool is not
+# the cache's supplier: the boot lands in vram_budget.cpp's `no lender` branch,
+# where the experts keep their plan-fixed 25.29 GiB and the cache gets the residue.
+# Measured twice — 12,288 and 49,152 MiB of host pool both printed the same
+# "kv cache 271.25 MiB ... 7,436"; only --gpu-headroom-mib moved anything
+# (768 -> 96 grew vram_weights 29.41 -> 30.06 GiB). The supplier the source names
+# is the arena: vram_budget.cpp:117, "lower --max-num-batched-tokens (the arena
+# scales with it)" — which is what the step size below now does.
+#
+# The pool, not the window, decides what the recorder accepts: it refuses any
+# document longer than the pool, so the 32,768-token window was decorative while the
+# cache held 7,436 — "E kld: document 196 is 8507 tokens and the KV pool holds
+# 7436, so it can never be scored" on all 197 records, exit 1. The longest record
+# is 10,554 tokens by the engine's own count (chars/3.6 for this file, not the
+# chars/4.97 this repo measured on the K3 corpus).
 #
 # The step size is ONE value for the recording and every scoring pass:
-# FRANK_QWEN38_MNBT, default 2,048. The gate's whole claim is that the reference,
+# FRANK_QWEN38_MNBT, default 1,024. The gate's whole claim is that the reference,
 # the served iq4r and every arm rode one shape, so the ceiling is set by the
 # TIGHTEST payload the gate boots rather than by the bf16 reference: the quantised
 # expert FILE tiers (the served iq4r, the arms) stream only at MNBT<=2048 and
-# refuse the widened form above it, and 2,048 is what production holds anyway.
-# Recording the reference at 32,768 would therefore measure a shape the payloads
-# cannot reproduce; at 2,048 the 50,432 positions cost more steps, which is still
-# one drained window. Calibration keeps the wide 65,536 — grams are not logits,
-# nothing downstream is compared against a step size, and that is where a big
-# step pays (a disk-tier step costs the same at 1 token as at 64k).
+# refuse the widened form above it. 2,048 was still the wrong default on this
+# payload, because the arena the engine charges for a step comes straight out of
+# the cache — 1.08 GiB at a 2,048-token step, 504.39 MiB at a 1,024-token one, and
+# the experts keep their share either way. At 1,024 the cache is 866 MiB: 24,300
+# tokens, which scores every one of the 197 documents (longest 11,970).
+# Calibration keeps the wide 65,536 — grams are not logits, nothing downstream is
+# compared against a step size, and that is where a big step pays.
 #
 # usage: scripts/run-qwen38-bf16.sh kldref|calib|score RAD OUT [--dry-run] [--image REF]
 #   (RAD and OUT are score's two positionals: RAD is a container under the models
@@ -86,8 +94,9 @@
 #        bf16 container, kldref only), FRANK_QWEN38_CAMPAIGN (holds kld-ref/ and
 #        calib/), FRANK_QWEN38_KLD_CORPUS, FRANK_QWEN38_TPL (the template copy
 #        the container reads through /models), FRANK_QWEN38_LOG_DIR, and
-#        FRANK_QWEN38_MNBT — the step size shared by kldref and score (2,048 is
-#        the file tier's ceiling and production's own value, four slots and all),
+#        FRANK_QWEN38_MNBT — the step size shared by kldref and score (default
+#        1,024: the file tier's ceiling is 2,048 and production runs there, but on
+#        this payload the arena charged for a wider step is taken from the cache),
 #        and FRANK_QWEN38_MAX_MODEL_LEN — the window every mode boots at (32,768;
 #        see the corpus census above, and raise it together with the corpus).
 # kldref refuses (exit 2) when the longest record of the corpus it is about to
@@ -148,12 +157,19 @@ KLD_REF_DIR="$CAMPAIGN/kld-ref"
 CALIB_DIR="$CAMPAIGN/calib"
 C=qwen38-bf16
 SERVICE=frank-archon
+# The card split, taken from the file that owns it instead of a private pair:
+# --host-pool-mib also funds the KV cache (see the header), so hardcoding it
+# here once starved the recorder. The `:?` guards keep an empty env fatal.
+HOST_POOL_MIB="${FRANK_ARCHON_HOST_POOL_MIB:?FRANK_ARCHON_HOST_POOL_MIB must be set}"
+GPU_HEADROOM_MIB="${FRANK_ARCHON_GPU_HEADROOM_MIB:?FRANK_ARCHON_GPU_HEADROOM_MIB must be set}"
 
 # ONE step size for both logits modes. The file tier streams only at <=2,048 and
 # the gate scores file-tier payloads (the served iq4r, the arms) against the
-# reference, so the reference is recorded at the ceiling the payloads can meet —
-# never two numbers that a KL between them would have to forgive.
-MNBT="${FRANK_QWEN38_MNBT:-2048}"
+# reference, so the reference is recorded at a step the payloads can meet — never
+# two numbers that a KL between them would have to forgive. 1,024 rather than the
+# 2,048 ceiling: on this payload the arena charged for the step is the cache's only
+# remaining supplier, and 1,024 is the rung that funds every corpus document.
+MNBT="${FRANK_QWEN38_MNBT:-1024}"
 if ! [[ "$MNBT" =~ ^[0-9]+$ ]]; then
     echo "REFUSED: FRANK_QWEN38_MNBT must be a token count (got: $MNBT)" >&2
     exit 2
@@ -201,7 +217,7 @@ fi
 # The measuring shape. Identical for kldref and score by construction — this
 # array is the only place the campaign's KL numbers are defined.
 ARGS=(--tp 2 --placement expert_tiered --weights-disk-tier
-      --host-pool-mib 12288 --gpu-headroom-mib 768
+      --host-pool-mib "$HOST_POOL_MIB" --gpu-headroom-mib "$GPU_HEADROOM_MIB"
       --deterministic --num-speculative-tokens 0
       --kv-cache-dtype fp8 --max-model-len "$MAX_MODEL_LEN"
       --ngram-placement ram
@@ -210,9 +226,10 @@ MOUNTS=(-v "$MODELS:/models:ro")
 TAPENV=()
 case "$MODE" in
   kldref)
-    # Recording is prefill-only over four slots and the shared step, which is
-    # production's own operating point (frank-archon.env: FRANK_ARCHON_MAX_SEQS=4
-    # with FRANK_ARCHON_MAX_BATCHED_TOKENS=2048). Corpus read-only, record dir rw.
+    # Recording is prefill-only over four slots at the campaign's single step size
+    # — 1,024, not production's 2,048 (frank-archon.env: FRANK_ARCHON_MAX_SEQS=4,
+    # FRANK_ARCHON_MAX_BATCHED_TOKENS=2048): the wider step costs the cache the
+    # room the corpus needs. Corpus read-only, record dir rw.
     ARGS+=(--model "$MODEL_IN"
            --kld-record /data/kld-ref --kld-corpus /data/kld.jsonl
            --max-num-seqs 4 --max-num-batched-tokens "$MNBT")

tests/qwen38_reference.bats

@@ -29,7 +29,7 @@ RUNNER="$REPO_ROOT/scripts/run-qwen38-bf16.sh"
 # The reference shape, spelled exactly as run-qwen38-bf16.sh must pass it: this
 # is the contract the recorded reference is measured at, so a silent edit here
 # is a silent edit to every KLD number in the campaign.
-SHAPE="--tp 2 --placement expert_tiered --weights-disk-tier --host-pool-mib 12288 --gpu-headroom-mib 768 --deterministic --num-speculative-tokens 0 --kv-cache-dtype fp8 --max-model-len 32768 --ngram-placement ram --override-chat-template /models/tpl.jinja"
+SHAPE="--tp 2 --placement expert_tiered --weights-disk-tier --host-pool-mib 49152 --gpu-headroom-mib 96 --deterministic --num-speculative-tokens 0 --kv-cache-dtype fp8 --max-model-len 32768 --ngram-placement ram --override-chat-template /models/tpl.jinja"
 
 # oneline — both dry runs print one shell-quoted argument a line; joining them
 # lets an option and its value be asserted as the pair they are.