fix(v4): the raw xs window may hold rows shorter than itself — lift of #116 onto master

[heydryft]

Lift of #116's two commits onto current master, unchanged. Same author, same content (73a0c72e3 + 6e408eb2e, cherry-picked); this exists because #116 is parked on #104 by branch topology, not by content, and the defect it fixes is real and measured.

#116 sits above #114 → #104, and #104 is held on a genuine open question — it and #122 (now master) are two mutually exclusive designs for ragged decode that contend on the clone_in_cache cadence, which needs an A/B nobody has run. None of that touches this fix.

Separability, proved rather than asserted

The two commits touch only kv_cache/xs_rolling.rs and models/deepseek4.rs. #104's work is scheduler/, engine/mod.rs, mtp_pipeline.rs, ragged_admission.rs. Disjoint. Cherry-picked onto 006657e05:

both applied clean
cargo check -p mistralrs-core --tests            -> exit 0   (unpiped)
cargo test  -p mistralrs-core --lib xs_rolling   -> 13 passed, 0 failed

including the three that carry the fix:

a_row_shorter_than_the_batched_window_is_accepted   ... ok
a_row_whose_promise_outruns_the_window_is_refused   ... ok
a_row_shorter_than_the_window_plans_and_never_reads_the_pad ... ok

The defect

ARC_V4_XS_PER_SEQ=1 ARC_MTP_PER_SEQ_KV=1 at B=8 returned 1 token, finish_reason: None, on 8/8 requests — measured twice on an H200 by #114, at be29f397, with zero occurrences in the OFF control and a full 48 tokens at B=1.

ERROR mistralrs_core::engine: completion step - Model failed with error:
  xs rolling cache: row 3 holds 9 tokens, fewer than the 11-wide retained window
  it would have to end at

Root cause, and it is a corrected diagnosis rather than the first theory. xs_rolling.rs refused any row whose token count was below the batched window's width. That is not the invariant the end-anchored layout needs — the file's own doc already stated the right one: base[i] >= tokens[i] - W, i.e. the row's promised run fits in the window. The code checked tokens >= W, which is sufficient to keep the offset arithmetic inside usize but not necessary for correctness, so it rejected rows that were perfectly representable.

So the fix is not the shape #114 ranked first. Every piece of the per-row machinery already existed: tokens/base are already Vec<usize>, clone_in_cache already front-pads a short row into the shared window (v_slack_at_front: true → pad_slack), and split_row already re-anchors on the way out. What was wrong was the invariant being checked.

What this does NOT include

#121's window pin (the measured +23.2% at uniform B=32) is deliberately not here, and it is not separable. Two independent reasons, both checked:

1. perf(v4): pin the xs window (+23.2% at uniform B=32) + the ladder that prices per-sequence advance — 4.46x at spread B=8 [MEASURED] #121's first commit 5ccd1a49a calls ragged_bucket_len(&running, **&self.admission**) in default_scheduler.rs — self.admission is feat(scheduler): admit ragged cohorts — the admission layer #103 named, and a general throughput bug #104's ragged-admission state. That commit cannot exist without feat(scheduler): admit ragged cohorts — the admission layer #103 named, and a general throughput bug #104, and every later perf(v4): pin the xs window (+23.2% at uniform B=32) + the ladder that prices per-sequence advance — 4.46x at spread B=8 [MEASURED] #121 commit sits on it.
2. Even the pin alone (57d330c8b, which touches only kv_cache/deepseek4 files) conflicts with this PR's own second commit: it removes compressor_needs_from(), which 6e408eb2e adds — 0 occurrences survive on perf(v4): pin the xs window (+23.2% at uniform B=32) + the ladder that prices per-sequence advance — 4.46x at spread B=8 [MEASURED] #121's tip. So fix(v4): the raw xs window may hold rows shorter than itself — B=8 per-seq advance generates again [MEASURED] #116→perf(v4): pin the xs window (+23.2% at uniform B=32) + the ladder that prices per-sequence advance — 4.46x at spread B=8 [MEASURED] #121 is a genuine rebase, not a set of independent cherry-picks.

⇒ #116 lifts; #121 does not. The pin stays parked until #104 resolves or its owner rebases the sequence deliberately.

Disposition of #116

#116 is not closed by this. Once this lands, its content is on master and it can close on its own terms, by its own author. Nothing has been force-pushed, rebased, or deleted — #116, #114, #121 and #104 are all untouched on the remote.

[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        24645        17702         4215         2728
 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                  143        14830        12217          797         1816
 Shell                    24         5924         4071         1457          396
 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                194        38699            0        29715         8984
 |- BASH                  71         1632         1192          324          116
 |- C                      3           17           17            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                  65         2048         1713           77          258
 |- TOML                   6          207          164            0           43
 |- YAML                   4           38           33            5            0
 (Total)                            44442         4671        30250         9521
─────────────────────────────────────────────────────────────────────────────────
 Rust                    663       313742       271458        14550        27734
 |- Markdown             478        24544          471        21171         2902
 (Total)                           338286       271929        35721        30636
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1284       461100       335289        76855        48956
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━