fix(BACKEND-TENSTORRENT-HOST-FREE-FORWARD): make captured decode reachable and token-clean at tip (#1105)

[lu-zero]

fix(BACKEND-TENSTORRENT-HOST-FREE-FORWARD): make captured decode reachable and token-clean at tip (#1105)

Four strands, one change: two capture-warming defects that kept every
capture-armed decode from completing at tip, one staging defect that
made the completing runs deterministically wrong, and one stale-state
defect that corrupted the first decoded token of every request after
the first in a multi-request capture session.

1. The capture-only embed segment (EmbedDeviceIdsInto and its
   hidden-shadow copy) had no eager predecessor, so its programs were cold
   at first capture and tt-metal refused to load binaries mid-trace
   (mesh_workload.cpp:153). The warm branch now runs the exact captured
   segment once before opening the scope; the hold semantics make the
   second run safe.

2. res.Zero at the top of the captured layer region left the fresh slot
   host-only: MemsetDeviceIfCapture refused a shadow-less slot and
   MemsetDeviceFill refuses under capture, so layer-0 EnsureDevice2D
   restaged from the recycled slot's stale persistent buffer — an
   enqueue_write, fatal at fd_mesh_command_queue.cpp:760. The no-shadow arm
   now serves the zero on-device under capture: in-place into the slot's
   persistent buffer when the [1, bytes/2] geometry matches, else a fresh
   ttnn::empty installed as persistent. Capture-only is load-bearing: bytes
   do not name a dtype (f32 KV masters share the pool blocks), so the eager
   fresh-slot zero keeps its host fallback.

3. The completing runs were still wrong: WarmPagedKvShadow described the
   flash-KV unbind(1) view — a rank-4 slice of the combined
   [nb,2,bs,nkv,d] store whose true block stride is 2bsnkv*d — as
   Tensor::Contiguous, so the shadow prefix upload indexed one slab early
   and, for every physical block >= 1, the K shadow received the V slab
   (zeros at prefill positions). The cold step computed on the mis-staged
   shadow (deterministic wrong first token, 48755 vs 3364) and replays
   re-consumed the same wrong-but-self-consistent state. The staging is now
   stride-aware: an IsFlashKvUnbindView helper, a general-stride upload
   path behind accept_unbind_view (defaults false, so every existing caller
   keeps strict contiguity), and the warm path passing the flag. The
   eager/default arm never entered the mis-staged path (its device PA
   declines on the same contiguity check and falls back to the host
   oracle), so it is untouched.

4. Even with clean staging, a recycled decode slot crossed a request
   boundary with a stale on-device position (TT capture arm: cross-request persistent KV-shadow state corrupts the second sequential request (capture battery prompt[1] tok=1) #2469):
   DecodePosCache()[num_reqs].cur_pos is a persistent device tensor aliased
   into both the RAC write index and the PA meta, and in the captured
   regime no steady-state path re-seeded it — WarmDecodePos early-returned
   on captured() and WarmPaMeta's r2_steady is process-global — so
   prompt[1] tok=1 attended prompt[0]'s decode KV rows still resident in
   the recycled physical block and RAC wrote at the stale virtual position;
   the wrong values decode to the previous prompt's own tokens. SizeSlot
   now tracks the position its trace expects and gates the replay regime on
   a per-slot continuation predicate; a same-size boundary routes through
   the TT host-free captured decode goes degenerate at the first KV block boundary; token-exactness evidence does not reproduce #1476 re-capture lane (destroy the trace, run the step eager against
   the freshly seeded position, re-capture on the next step) — the boundary
   seed alone measured insufficient and, without the lane, the battery does
   not complete at all. WarmPaMeta now echoes a device-truth mirror into
   cp_host, so the device-PA guard reads what the device holds instead of
   comparing the host's expectation against itself (the repaired guard and
   its host fallback jointly absorb a missed seed, so the battery pins the
   pair, not the predicate alone).

Evidence (P150, one GPU lock, fresh reviewer mutations all red with
sha256-verified restoration): focused capture-armed gate token-clean at
the cold step (3364 == default arm, coherent text, 0 fatals, healthy
replay count); at the merged head the capture-armed golden battery
passes 125/125 assertions and 16/16 prompts with a max top-2 gap of
0.375 nats at the formerly-failing cell (prompt[1] tok=1, inside the
row's <=0.5 near-tie band), the default arm identical, and the capture
CLI byte-identical across three runs including one after a card reset;
same-binary A/B/C warm medians default 12.90 / opt-out 17.80 / captured
27.47 tok/s (~2.1x / ~1.6x), the capture legs coherent for the first
time. Qwen3-4B showed the same capture-arm corruption before the fixes
and is covered by the same backend-level repairs. The earlier wave
record's "coherent" claim for a pre-repair head was a measurement error
(a word-count gate counted tt-metal log lines); the record correction
(3d06b42) is part of this change.

The capture arm stays opt-in (VT_TT_DECODE_CAPTURE). The capture-default
flip stays blocked on #1625 (cross-prompt geometry re-warm spins in pure
userspace under a live trace); the multi-request exactness blocker
#2469 is fixed by strand 4, with the continuous-batching (row r>0) scope
and the uncounted guard-decline counter recorded under the spec's

Owed.

Closes #2461
Closes #2469

FOLLOWING_AGENTS_PROTOCOL

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