polyglot-rust-c × 3 at budget 8000 — 0/3 (expected-fail task), but two clear signals: B2 oversized writes + a NEW soft-runaway variant

What was run

polyglot-rust-c × 3 at stock settings, budget 8000. This is the suite's canonical hard task (historically failed as a "B1 looper"); the question was whether the bigger budget changes its failure mode.

Results — all three fail, none for the old reason

• FQfZzxS (FAIL, 4m24s, 5 turns) — the new finding: a soft runaway. Every turn's thinking maxes the budget (26.6k chars ≈ 7.6k tok — polyglot saturates any budget). On the final turn the cut reasoning continued as 19k chars of plain TEXT ("Wait, I need to reconsider…"), wandered, and self-terminated with a clean stop — its tail repeats the thinking tail verbatim. No tool call → pi's print loop ends "normally" → trial dies half-done. Crucially this evades the runaway recovery, which only armed on length stops.
• Nox4NEM (FAIL, 25m43s, 19 turns): 3 length-stops with tool calls = B2 oversized writes (tool args up to 89.5k chars). Guard engaged sanely (pr:2 — two batched advances, vs 79 breaks in the per-call era).
• dhpwZQW (ERR AgentTimeoutError, 30m, 7 turns): 5 length-stops, tool args up to 82.8k chars — pure B2. Peak context 147k, no prunes needed.

The recovery correctly stayed silent on the B2 trials (toolCall present) and had no trigger for FQfZzxS (clean stop). Zero false fires; zero true fires.

Actions taken

1. Recovery trigger extended in harnesses/minimal_pi.py: now also fires on stopReason "stop" + no toolCall + text > 8000 chars (largest legitimate final answer observed anywhere today: 1.9k chars; artifact range 19–75k). Same trim + followUp nudge, same 2/session cap. 16/16 unit tests incl. three new soft-runaway cases; hysteresis suite still 15/15.
2. B2 oversized writes remain the task's real blocker → H2 chunked-write scaffolding is the next lever here. Do NOT lower max_tokens.

Verdict on the budget sweep

Keep 8000: it fixed the artifact where thinking fits (regex-log 6/6 today) and polyglot shows no budget value would fit its thinking anyway — the residual is a harness-recovery problem (now addressed, pending live proof) plus a B2 problem (H2).