Hysteresis smoke — guard correct in 5/5; both fails are the SAME artifact: reasoning-budget force-close → plain-text thinking continuation → 32k runaway

What was run

regex-log × 5, COMPACT_AT_TOKENS=15000, first run with frontier hysteresis (stable stub set between calls, batched advances) + live ratio calibration.

Guard verdict: working as designed

• Advances are batched with full logging (frontier N/M): 5–16 advances per pruning trial instead of pruning every call; stub sets stayed stable between advances; est held ≤ target after every advance.
• Live calibration visibly at work: ratio climbs 1.7 → 2.26/2.45 within trials as dense content accumulates (no more freeze).
• The 3 non-looping pruned trials all passed (2m50s–10m); peaks 14–39k real. No guard-attributable failure: one fail looped at turn 1 with zero prunes, the other looped after 42 productive pruned turns while three equally-pruned trials passed.

The two failures — smoking-gun evidence on the budget force-close

Both are length-stop plain-text runaways (B1), and both show the same mechanism:

• xqqcPFL (turn 1, no tools): thinking cut mid-regex, inside a character class (…(?:25[0-5]|2[) by the server's --reasoning-budget 4000 force-close — and the TEXT block begins with the character-for- character continuation of that same regex (]|1[0-9]{2}|…), then walks matches for 75k chars until the 32k output cap. The model never noticed it left <think>; its reasoning simply continued in plain text, uncapped.
• hScZhud (turn 43, after 42 good turns): same shape — thinking ends mid-trace-walk, text resumes the analysis ("I see the issue…") and spirals for 63k chars to the cap. The length-stop has no tool call, so the run ends with the regex unfixed.
• tBaJ2he passed despite a 62k-char text ramble mid-run — the same artifact, survived by luck.

So: the budget force-close doesn't just cause degenerate stops (the > char, the user leak) — it causes plain-text thinking-continuation runaways, and tonight it cost 2/5 trials on a task with a ~90% historical pass rate. THINK sizes at the cut: 9–13k chars ≈ 2.9–3.8k tokens, right at the 4000 budget. This mechanistically links two planned experiments: 1. Reasoning-budget sweep / --reasoning-budget-message tuning (server-side) — attack the trigger; 2. DRY sampler — attack the symptom (the loop itself).

Cache A/B at limit 15000: neutral, as expected in hindsight

Hit rate 65.4% (hysteresis) vs 73.1% (per-call pruning, run …150817) — a wash, confounded by the two looper trials (43/46-request trajectories, 60k+ char text blocks). The structural reason: at limit 15000 the limit−target slack is only ~3.7k real-equivalent tokens ≈ 1–2 turns, so the frontier advances almost every few calls anyway. At the production limit (200000, slack 50k ≈ 15–30 turns per advance) the batching actually pays. Decisive measurement = rerunning make-mips-interpreter at stock settings (baseline: 62.4% hit, 79 cache breaks, ~23 min re-prefill).

Server note: no --cache-reuse in the llama-swap config — strict prefix caching, so the stable-prefix property is the only cache lever available.

Next

1. make smoke MODELS=qwen3.6-35b-a3b SMOKE_TASKS=make-mips-interpreter (~1h) — the real cache measurement: expect single-digit pr: (vs 79) and hit rate back toward ~90%.
2. Reasoning-budget sweep 2000/4000/8000 + budget-message phrasing — now the top score lever (NOT during a live run; llama-swap hot-reloads).
3. DRY sampler A/B (polyglot-rust-c should improve, circuit-fibsqrt unchanged).