[noahgift]

K1: apr serve sizes the GPU KV from the model's context (#4603), and refuses past it with a 400 (#4614)

Base: main. This PR carries only the KV fix and its test. The rest of the earlier stack (#3715 serial prefill / for_session, the #3413 message, #4620, #4623, #4624) stays on its own branches for a later PR. Nothing was deleted.

Rows (3)

1. fix(#3715 B1): release-readiness ENFORCES at T-1/T-4; report arm deleted (R10) #4603 K1. apr serve builds the CUDA model with OwnedQuantizedModelCuda::for_serving. The device KV now equals the model's context, capped by free VRAM (with 2 batched slots budgeted), and never drops below the old 4096. Before this change, serve fixed the KV at 4096 whatever the model declared, so a longer prompt overflowed the cache and returned a 500. --context-length still overrides.
2. fix(serve): a turn past the device KV cache is a 400 "context exceeds N", not a 500 (D5) #4614 / PMAT-4616 D5. A prompt at or past the serving context is refused whole with 400 context exceeds N … refused whole, not truncated on chat, completions and stream. It is never a 500 and never a truncated 200.
3. KV budget core: gguf/cuda/session_kv.rs and gguf/cuda/resident.rs, which for_serving sizes from.

Test

api/tests/serve_kv_limit_k1.rs (cuda-only, needs no model file, so it runs in cuda-unit) builds a synthetic 8192-context model. It serves the model through the real router and batch scheduler on the device, then asserts:

• the KV is 8192;
• a prompt of 8188 plus 4 generated tokens returns 200 with all 4 tokens;
• prompts of 8192 and 8193 return 400 "refused whole".

Reverting to a fixed 4096 fails the KV assertion.

Dev run on gx10 (not CI evidence):

• The aprender-serve --features cuda filter (k1, d5, session_kv, resident) passed 28/28.
• The apr-cli serve tests passed 429/429.
• The revert to Fixed(4096) turned the test red.

Refs #4603 #4614

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=4630 head=da4f4f9c627866f17a3ce5f7f65064fc04d7fbb4 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4616",
 "head": "6a6a1c2de8fe92b4ff4947ff20b048f905c362bc",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 1
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 1
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

quorum-review (AD-04): three PASS — agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-4616,PMAT-4636",
 "head": "45d616e6e75802b6f5761c81ca0cb402e62c3d86",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 1
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

Folded into #4623 at 616ff4c; branch kept. The fold is the union of both diffs against main, so this PR's serve KV sizing and #4623's chat session KV both stay.