pillar1(R8): isolate long-seq bug to batched prefill path (partial)

Extended TQ_DUMP_HIDDEN to support TQ_DUMP_POS={0|N|all} for per-
position dumps (infrastructure for the next-step per-layer diff).

A/B on Qwen3-0.6B with 144-token synthetic input:
  batched prefill:   UTF-8 garbage (definitively broken)
  per-token prefill: ASCII but still wrong (secondary issue)
  KV fp32 (both):    same patterns — KV compression NOT the cause

Primary follow-on target: tq_forward_batch non-MoE path. Per-token
path has a separate, milder issue (likely RoPE/attention at larger
positions). Both need the HF per-layer diff to pinpoint.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: unamedkr

bench/results/2026-04-20_longseq_transformer_bug.md

@@ -56,6 +56,28 @@ Confirmed: our Qwen3-0.6B produces token 9707 for "Hello" (matching HF).
 4. **Attention dispatch branch at seq_len > 128 / > 256**
    - Some kernels have special-case paths for long sequences.
 
+## R8 partial isolation: batched prefill is the primary offender
+
+A/B on Qwen3-0.6B with 50-synthetic-word prompt (144 tokens):
+
+| Path | Output |
+|---|---|
+| Batched (`tq_forward_batch`, default) | `alyticsанciea��...` (UTF-8 garbage) |
+| Per-token (`TQ_NO_BATCH_PREFILL=1`) | `" =, on up = a,="` (ASCII, broken but less) |
+| KV fp32 + batched | `���isonswana...` (still garbage, KV quant not root cause) |
+| KV fp32 + per-token | same as per-token above |
+
+Interpretation:
+- Batched path is definitively broken — produces pure UTF-8 byte garbage.
+- Per-token path is also producing wrong output on natural prose, but
+  with ASCII characters rather than byte-level chaos. Some subtle
+  accumulation issue separate from the batched bug.
+- KV compression is NOT the cause; fp32 KV shows identical pattern.
+
+Primary target for follow-on: `tq_forward_batch` for non-MoE models.
+Secondary: per-token path on natural prose (may be RoPE / attention
+accumulation at larger pos).
+
 ## Next steps (methodology)
 
 Apply Pillar 1 methodology to transformer forward:

src/engine/tq_transformer.c

@@ -2489,11 +2489,23 @@ static void self_attn_forward(tq_model_t* model, tq_state_t* s, int l, int pos)
  * Dumps only at pos=0 (first token of prefill/generation) to avoid
  * overwriting across prefill tokens. */
 static void tq_dump_hidden(const char* name, const float* data, int n, int pos) {
-    if (pos != 0) return;
     const char* dir = getenv("TQ_DUMP_HIDDEN");
     if (!dir) return;
+    /* TQ_DUMP_POS=N selects a single position; default 0 preserves old
+     * behavior. TQ_DUMP_POS=all dumps every position (expensive: 28 × N
+     * files per forward pass). */
+    const char* pos_env = getenv("TQ_DUMP_POS");
+    if (pos_env && strcmp(pos_env, "all") == 0) {
+        /* dump all — append position to filename */
+    } else {
+        int target = pos_env ? atoi(pos_env) : 0;
+        if (pos != target) return;
+    }
     char path[512];
-    snprintf(path, sizeof(path), "%s/%s.bin", dir, name);
+    if (pos_env && strcmp(pos_env, "all") == 0)
+        snprintf(path, sizeof(path), "%s/%s_p%d.bin", dir, name, pos);
+    else
+        snprintf(path, sizeof(path), "%s/%s.bin", dir, name);
     FILE* f = fopen(path, "wb");
     if (!f) return;
     fwrite(data, sizeof(float), (size_t)n, f);