fix(v4): close the audit's last open finding-4 item; the long-prompt death is not a reference divergence

[heydryft]

Closes the reference-audit follow-up (PR #27, task #16) and answers the
long-prompt serving failure (task #22).

The six findings were already fixed

All of PR #27's two CRITICAL and four HIGH findings are on master. Verified by
reading the current code, not by trusting the PR titles:

#	Finding	State
1	CRITICAL — MTP draft seeded with embed(T0), not the target's hidden state	fixed (#30) — propose_chain is seeded from last_hidden; v4_mtp_draft_depends_on_the_target_hidden_state_not_only_the_token fails on the old seeding
2	CRITICAL — MTP attends an empty per-chain KV at absolute RoPE positions	fixed (#30) — new_draft_cache is persistent per sequence and extended over accepted tokens (extend_draft_kv_row), mirroring forward_draft_extend; pinned by v4_mtp_draft_kv_holds_the_accepted_context
3	HIGH — b_sz > 1 prefill routed to mask-free sinks_attn_varlen	fixed (#28) — is_varlen now requires mask.is_none(), and dsv4_attention always passes Some(&mask), so V4 can never reach it
4	HIGH — caller's attention_mask discarded	fixed (#28) — compose_caller_mask folds it in and errors on a width it cannot place. One sub-item was left open and is fixed in this PR (below)
5	HIGH — swiglu clamp missing on 4 of 5 expert paths	fixed (#35) — every backend routes activation through swiglu_clamp; the shared expert carries .with_swiglu_limit(cfg.swiglu_limit)
6	HIGH — YaRN applied to Standard (ratio-0) layers	fixed (#35) — rope_scaling_for_compress_ratio returns None at ratio 0; the old behaviour is behind ARC_YARN_ON_STANDARD_LAYERS=1

What this PR changes

SdpaParams::sliding_window was inverted (audit §1(e), the one open piece of
finding 4). sinks_attn forwards it to the fused flash-sinks kernels as
window_size — a bottom-right-aligned relative-distance window over whatever
key axis those kernels are handed. dsv4_attention hands them the raw sliding
window on Standard layers and the concatenation [raw ++ compressed] on
CSA/HCA. So Standard is the only ratio that may declare a window, and it was the
one declaring None. On CSA/HCA a window over the union axis deletes exactly the
distant-context keys the compressed branch exists to supply — compressed entry
j stands for absolute position j * ratio, not t_k - j.

Inert today only because flash_sinks_ok rejects head_dim = 512, the head dim
every V4 layer uses. PR #94 adds 512-wide kernels, which arms it — silently,
because the fused path takes no mask that could disagree with it.

Length coverage. v4_long_prompt_prefill_and_decode_sweep prefills and then
decodes through the real V4 dispatch at 127/129/511/513/1400 tokens, bracketing
the 128-token window, the CSA/HCA block boundaries, and CACHE_GROW_SIZE (512).
CPU unit test, labelled as such — not a GPU claim.

Both new tests are mutation-checked: inverting sdpa_sliding_window fails the
first; pinning XsRollingCache's row ceiling to its initial 64 fails the second
at T=511 and only there, which is what makes the sweep length-selective rather
than a smoke test.

Does any finding explain the long-prompt death?

No. All six are eliminated, and the mechanism is throughput, not correctness.

The measurement (FACTS, session 8): 9 words → 100.15 s for 8 tokens; 198 words →
386.15 s for 8 tokens; ~1,055 words → connection refused.

Elimination:

• Findings 1, 2, 5, 6 carry no length dependence at all — MTP seeding, the
  clamp and the RoPE table are identical at 12 tokens and at 1,400.
• Finding 3 needs b_sz > 1; a single canary request is b = 1, and the
  route is closed regardless.
• Finding 4 is the only length-shaped candidate, and the sweep above is the
  direct test: prefill + decode at 1,400 tokens through the real dispatch is
  clean and finite. The shape/mask/position/compressed-KV class — a mask built
  wrong, a position index that mis-bases, a compressed-KV path that only holds
  short — is ruled out on CPU.
• A panic cannot do it either: step_catching_panics turns an engine panic into
  a failed batch, and handle_pipeline_forward_error! turns an Err into a
  failed request. Neither closes the socket. "Connection refused" means the
  process left.

What is left is cost. V4 has no fused attention kernel at head_dim = 512
(flash_sinks_ok = {64,80,96,112,128,192,256}), so every prefill layer runs
unfused through sinks_attn_cpu, which repeat_kv-expands K and V to all 64
heads with a real Tensor::cat (attention/mod.rs:98) and materialises the full
score matrix. At the real geometry, per layer, bf16, CSA (n_keys = 1.25·T):

• repeat_kv K+V: 2 · 64 · 1.25T · 512 · 2 B = 163,840·T bytes
• scores, twice (matmul out + softmax out): 2 · 64 · 1.25T² · 2 B = 320·T² bytes

⇒ 71 MB/layer at T = 280, 856 MB/layer at T = 1400, 5.6 GB/layer at T = 4096,
against 43 layers of arithmetic that is 7.05e6·T² FLOP for the stack.

Fitting the two OK cells to t = a + b·T² (≈1.33 tok/word) gives a ≈ 96 s,
b ≈ 4.3e-3 s/token² — and predicts ≈ 8,500 s (2.4 h) for the 1,055-word
prompt. A request that never returns is indistinguishable from a dead server to
a canary, and the transient working set is an order of magnitude larger at the
same time.

Stated honestly: two points fit a two-parameter curve exactly, so this is
consistent with quadratic prefill, not a measurement of it. The a ≈ 96 s
length-independent term is separately corroborated — it is 8 decode steps of the
unfused decode path plus the CPU sampler fallback already logged in FACTS
(tensor_device_ptr: unsupported dtype I32, ~10/sec), which another agent owns.
That fallback is additive to a and does not touch b, so fixing it will not
move the long-prompt cell.

⇒ Task #22 needs a separate owner, and the owner is the kernel workstream, not
the audit. The two things that would settle it on a box, in order:

1. one prefill-only timing at T ∈ {256, 512, 1024} — three points test the
   quadratic rather than assuming it, and separate a from b;
2. nvidia-smi peak reserved during the 1,055-word attempt, to say whether it
   died of the clock or of memory.

Both fit inside any rental already up.

Also surfaced, not shipped

• The CUDA-graph decode arm (deepseek4.rs, has_graph_mode_positions())
  still passes raw_prefix: 0 while reading a fixed cap = sliding_window-wide
  rolling window, so q0 is pinned at cap - 1 and the compressed-block
  threshold freezes at floor(128/ratio) — the distant-context branch stops
  growing past token 128. This is the second half of audit §1(e) note 2. Not
  fixed here: the correct raw_prefix is a host value and the position is
  deliberately device-side, so it needs a design decision (plumb the host
  position, or refuse compressed layers on the graph arm), not a patch. It is
  behind a double opt-in (ARC_V4_CAPTURE_PROBE + ARC_CANDLE_ALLOC_CACHE) and
  is not on the serving path.
• MtpSpeculativePipeline::extend_draft_kv is dead (extend_draft_kv_row is the
  live one). Pre-existing.

🤖 Generated with Claude Code

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                  5          305          210           52           43
 CSS                       2         1181         1036           34          111
 CUDA                     72        24328        17592         4018         2718
 Dockerfile                1           39           22            8            9
 JavaScript               16         3546         2676          482          388
 Jinja2                    7          694          656            5           33
 JSON                     74         4600         4597            0            3
 Makefile                  1            6            5            0            1
 Metal Shading Lan|       33        12224         9431         1142         1651
 PowerShell                1          300          227           30           43
 Python                  140        14489        11930          782         1777
 Shell                    21         4531         3182         1056          293
 Plain Text                4         3801            0         2479         1322
 TOML                     33         1485         1292           43          150
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 HTML                      4         2687         2604           43           40
 |- CSS                    2          543          479           37           27
 |- JavaScript             1         1233         1215           12            6
 (Total)                             4463         4298           92           73
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         4          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                182        35380            0        27061         8319
 |- BASH                  65         1567         1163          294          110
 |- C                      1           10           10            0            0
 |- CUDA                   2           84           56           16           12
 |- JSON                  18          708          708            0            0
 |- PowerShell             1            1            1            0            0
 |- Python                23         1008          787          113          108
 |- Rust                  64         2043         1710           75          258
 |- TOML                   6          207          164            0           43
 |- YAML                   4           38           33            5            0
 (Total)                            41046         4632        27564         8850
─────────────────────────────────────────────────────────────────────────────────
 Rust                    660       305484       264599        13728        27157
 |- Markdown             475        22019          471        18941         2607
 (Total)                           327503       265070        32669        29764
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1263       444870       327105        70504        47261
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[heydryft]

Admin-merged without review (this fork has no second reviewer); all 16 checks green at merge time — verified against head ef9dfce9, 0 pending, 0 failures. Merged deliberately ahead of #94: the inverted SdpaParams::sliding_window this fixes is inert today only because flash_sinks_ok rejects head_dim 512, and #94's 512-wide kernels arm it.