Cache length after prefill (P=3): 3

Running 5 decode steps with cached-vs-uncached parity checks:

 step  cache_len    max|diff|  parity_ok
---------------------------------------------
    1          4    1.788e-07       True
    2          5    1.788e-07       True
    3          6    1.192e-07       True
    4          7    2.384e-07       True
    5          8    2.384e-07       True
---------------------------------------------
All 5 steps within tolerance 1e-05: True

PASS: cached and uncached next-token logits agree for every decode step.
Cache length after prefill (P=3): 3

Per-step cache growth check (4 decode steps):

 step   expected     actual    per_layer     ok
-------------------------------------------------
    1          4          4         True   True
    2          5          5         True   True
    3          6          6         True   True
    4          7          7         True   True
-------------------------------------------------
Final cache length: 7 (expected 7)
All layers grow by exactly one column per step: True

PASS: one KV append per layer per step; cache shape [B, H_kv, P+S, D_HEAD] confirmed.
Raw KV-cache bytes from the lab model dimensions

  L (layers)     = 2
  H_q (Q heads)  = 2
  D_HEAD         = 16
  S (cached tok) = 64
  B (batch)      = 1
  b (bytes/elem) = 2 (FP16)

   Variant   H_kv    Per-layer bytes   Total (2 layers)   Ratio vs MHA
----------------------------------------------------------------------
       MHA      2              8,192             16,384           1.00
       GQA      1              4,096              8,192           0.50
----------------------------------------------------------------------

MHA: 8,192 bytes per layer; 16,384 bytes total across 2 layers
GQA: 4,096 bytes per layer; 8,192 bytes total across 2 layers
GQA figures are calculated from H_kv=1, not measured from an instantiated cache.
GQA / MHA ratio:     0.50 (50% reduction)

Note: cache-size ratios only. Total model memory is dominated by weights,
and latency depends on kernel fusion, bandwidth, and compute vs memory bound.
This is raw tensor bytes, not GPU allocator memory.

PASS: per-layer and total byte figures are consistent.
============================================================
BROKEN cache-position case (new K/V overwrites slot 0)
============================================================
  step 1: max|cached - uncached| = 1.654e-01  (tolerance 1e-05) -> MISMATCH (expected)
  step 2: max|cached - uncached| = 2.335e-01  (tolerance 1e-05) -> MISMATCH (expected)
  step 3: max|cached - uncached| = 2.875e-01  (tolerance 1e-05) -> MISMATCH (expected)

  All 3 steps show mismatch (expected for broken cache): True

============================================================
FIXED cache-position case (correct write at write_position)
============================================================
  step 1: max|cached - uncached| = 2.384e-07  (tolerance 1e-05) -> OK
  step 2: max|cached - uncached| = 2.384e-07  (tolerance 1e-05) -> OK
  step 3: max|cached - uncached| = 2.384e-07  (tolerance 1e-05) -> OK

  All 3 steps within tolerance: True

============================================================
DIAGNOSIS AND REPAIR
============================================================
Bug: Each new K/V pair was written to cache slot 0 instead of being
     appended at write_position (the current cache length). The cache
     never grew, and the newest token overwrote the oldest entry, so
     later reads attended to a corrupted key/value history.

Fix: Write new K/V at slot `k_cache.shape[2]` (exactly the current
     cache length), i.e. the next free position. After the fix, cached
     and uncached logits agree within FP32 tolerance.

PASS: broken cache-position detected and repaired.