feat(BACKEND-TENSTORRENT-KEEPQUANT): the 27B fits — packed GDN projections, int8-dot captured decode, per-request logits gather (#3042)

[lu-zero]

feat(BACKEND-TENSTORRENT-KEEPQUANT): the 27B fits — packed GDN projections, int8-dot captured decode, per-request logits gather (#3042)

The Qwen3.8-27B Q4_K_M token gate is GREEN on the Blackhole P150:
16/16 prompts, 0 forward-divergent, max gap 0.062 nats against the
llama.cpp-b10451 oracle (band 500 mnats), knob-free captured decode.
Backend proof: device type 6, 0 declines — kPagedAttention 4096,
kGdnDecode 11520, kCausalConv1dUpdate 11520 selections. The founding
blocker of the keep-quant row — GGUF k-quant arms refuse on the P150 —
is resolved end to end.

Three stacked defects stood between the card and the verdict, each
root-caused with the W4d W0 allocation ledger and fixed here:

1. THE PLACEMENT PROBE (landed in fix(BACKEND-TENSTORRENT-KEEPQUANT): probe the TT DRAM total so the device fit can resolve (#3042) #3165): the TT platform reported no
   device memory total, so the device fit refused every qwen3.8
   checkpoint and placed them on CPU — the card was never reached.

2. THE 12 GIB BF16 GDN RESIDENCY: the V-head reorder routed the GDN
   projection family (attn_qkv q6_K, attn_gate q4_K) to
   kTransformedWeight -> kExpandBf16 — 97 x [10240,5120] bf16 stagings,
   9.7 GiB beside the packed words. The reorder permutes WHOLE rows, so
   the fix permutes the packed bytes (ReorderVRows<uint8_t> — the
   template is type-agnostic) and keeps the projections block-encoded:
   the GDN dispatch decodes them through their keep-quant words
   (MatmulBTRawD over the split-packed pair; the merged arm needs one
   encoding, this checkpoint mixes q6_K qkv with a q4_K gate). Slot
   residency 23.8 -> 7.7 GiB, measured; all 64 blocks warm (previously
   died at block 55). Equivalence proven bit-for-bit: packed-byte
   reorder == element reorder through the real dequantizer, q6_K +
   q4_K (standalone TT-free check; ASan also caught a row_off-scaling
   bug in the CHECK'S OWN reference — the shipped template was
   correct).

3. THE CAPTURE-TIME TRANSIENTS: the E=1 chunk-decode chain's planes are
   pinned into the trace region for replay (>= 3.8 GiB at 27B —
   mesh_trace.cpp:82 — invariant to chunk size), and the lm_head's
   [31040,5120] dequant chunk (606 MiB) fragmented the banks besides.
   W4b's int8-dot kernel was built for exactly this capture-demand gate
   and this wave turns it on for every captured keep-quant matmul: the
   bf16-out arms join via an explicit f32->bf16 cast before the commit
   (the W4b store-geometry bug — f32 bytes in a bf16 slot — fixed and
   red/green-proven by a device-consumer probe: the uncast output fed a
   second device matmul produces non-finite garbage). Zero kq-decode
   labels fire inside the capture scope now.

Plus: the keep-quant word shadows stage at LOAD (a lazy first-matmul
stage at 27B asked the head's 1 GiB words against a 100 MB largest free
block); the per-request logits gather lands (vLLM logits_processor
semantics, LastTokenLogitsIndices, both driver Step fallbacks — the
mixed-step lm_head computed all T rows of logits when generation samples
only the last token per sequence); and the 0.8B vehicle keeps its
committed arms untouched (the int8-dot lever is env-gated
VT_TT_KEEPQUANT_INT8DOT, default off).

Goldens: the anchor pair derives from the engine's own bootstrap
capture, teacher-forced through the pinned llama.cpp b10451 oracle
(DequantGgufRowToF32-verified bit-for-bit), and the GREEN run validates
it as this arm's grading anchor: the device ids landed inside the
near-tie band at every cell.

The 0.8B vehicle gate: 16/16, byte-identical to its committed baseline
(STRICT 8/16, near-tie 8/16, max 0.25 nats). Backend suite 73/73
(524,442 assertions) including the four bit-exact keep-quant sweeps.
The three specs on this branch (W4 packed reorder, W5 gather — its
trace-region premise falsified by the trace and amended in-file, W6
int8-dot capture) carry the full evidence chain.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]