feat(intern-decision): persist the triton autotune cache across recreates

The ~6.5-9 s first-call-per-bucket autotune lived in the container's writable
layer and died on every recreate. 0.1.3 creates /tmp/triton-cache in the image
owned by 10001 so the named volume intern-decision_triton-cache inherits a
writable mount point, and compose mounts it.

Bucket model PROVEN, not inferred: 2,048-token buckets, 16 up to 32,768. After
one warmed call per bucket, 12 random sizes across 8k-32k were all warm (worst
2.09 s); cold entries cost 6.5-9 s. Full cold warm-up 109 s; warm re-run 17 s.

scripts/intern-decision-warmup: one noul call per bucket, MAX_TOKENS from
/health, two-point live calibration of the tokenizer's linear token model (a
single probe overcorrects and the aim oscillates around the bucket edge),
per-bucket wall times, non-zero exit on a missed bucket. Run it after an IMAGE
CHANGE only; the volume carries ordinary recreates (measured: force-recreate,
then a warmed 32k call answered in 2.11 s).

Acceptance on 0.1.3: JevBench 202/231, hard 83/111, 0 diffs / 924; warm 32k GPU
1 peak 15,218 MiB (budget 15,220; a COLD autotune touched 15,224 once, README
caveat); /decide answers. Artifacts in the acceptance dir.

scripts/intern-decision-warmup

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Warm every length bucket of intern-decision's kernel autotune cache.
+
+The fast kernels (triton + fla) autotune per shape BUCKET: the first call that lands in a new
+bucket costs seconds of autotune. The cache lives in the container's `triton-cache` volume, so
+it survives ordinary recreates — this script is for AFTER AN IMAGE CHANGE (a new triton/fla
+version keys new entries). Idempotent: a second run is all fast.
+
+One noul /v1/systemone call per bucket, smallest up to the service's own MAX_TOKENS (read from
+/health; never hard-coded). Aims just under each bucket's top edge; the aim is calibrated from
+the service's own reported input_tokens after every call, so drift cannot leave a bucket cold.
+Prints each call's bucket, the model's own token count, and the wall time. Exits non-zero if
+any call fails or lands outside its bucket.
+
+    scripts/intern-decision-warmup [--endpoint URL] [--bucket-width N]
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+
+UNIT = "the quick brown fox. "        # ~5.13 tokens against this service (measured 2026-09-30)
+
+
+def post(endpoint: str, token: str, body: dict, timeout: float = 180.0):
+    req = urllib.request.Request(
+        endpoint.rstrip("/") + "/v1/systemone",
+        data=json.dumps(body).encode(), method="POST",
+        headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"})
+    t0 = time.perf_counter()
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as r:
+            return r.status, json.loads(r.read()), time.perf_counter() - t0
+    except urllib.error.HTTPError as e:
+        return e.code, json.loads(e.read() or b"{}"), time.perf_counter() - t0
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--endpoint", default="http://intern-decision.fv.internal:8033")
+    ap.add_argument("--bucket-width", type=int, default=2048,
+                    help="asserted bucket width; the acceptance run proves or corrects it")
+    args = ap.parse_args()
+
+    token = subprocess.run(["secret", "get", "intern-decision/api-token"],
+                           capture_output=True, text=True).stdout.strip()
+    if not token:
+        print("no token: secret get intern-decision/api-token", file=sys.stderr)
+        return 2
+    health = json.load(urllib.request.urlopen(args.endpoint.rstrip("/") + "/health", timeout=15))
+    max_tokens = int(health["max_tokens"])
+    width = args.bucket_width
+    buckets = list(range(width, max_tokens + 1, width))
+    print(f"endpoint={args.endpoint} max_tokens={max_tokens} bucket_width={width} "
+          f"buckets={len(buckets)}")
+
+    def ask(reps: int):
+        return post(args.endpoint, token, {"state": UNIT * reps,
+                       "questions": {"warm": {"type": "noul",
+                                              "instructions": "Is the system healthy?"}}})
+
+    # Two-point calibration of the tokenizer's linear model, input_tokens = a*reps + b. A single
+    # probe cannot separate a from b and its "calibration" overcorrects on the next aim (measured:
+    # the aim oscillates ±500 tokens around the bucket edge). Two probes 900 reps apart fit exactly.
+    s1, o1, _ = ask(100)
+    s2, o2, _ = ask(1000)
+    if s1 != 200 or s2 != 200:
+        print("calibration failed", s1, s2, file=sys.stderr)
+        return 2
+    t1 = o1["usage"]["input_tokens"]; t2 = o2["usage"]["input_tokens"]
+    a = (t2 - t1) / 900.0
+    b = t1 - a * 100
+    print(f"calibrated: input_tokens = {a:.4f}*reps + {b:.1f}")
+
+    failed, total = False, 0.0
+    print(f"{'bucket<=':>9}  {'tokens':>7}  {'wall_s':>7}  note")
+    for hi in buckets:
+        note = ""
+        got = status = 0
+        wall = "  n/a"
+        for attempt in (1, 2):           # a second aim fixes a rounding miss; never miss a bucket
+            reps = max(1, round((hi - 60 - b) / a))
+            status, out, w = ask(reps)
+            total += w
+            wall = f"{w:.2f}"
+            got = (out.get("usage") or {}).get("input_tokens", 0)
+            if status != 200 or hi - width < got <= hi:
+                break
+        if status != 200:
+            note = f"FAILED {status}: {(out.get('error') or {}).get('message', '')[:60]}"
+            failed = True
+        elif not (hi - width < got <= hi):
+            note = f"OUT OF BUCKET ({hi-width} < {got} <= {hi})"
+            failed = True
+        print(f"{hi:>9}  {got:>7}  {wall:>7}  {note}")
+    print(f"total {total:.1f} s over {len(buckets)} buckets")
+    return 1 if failed else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())