[Air000000]

Changes

• Explicitly disable thinking for Qwen3.5 Plus generation, including dated variants; do not inject the vendor option into unrelated configured models.
• Preserve the 30-second timeout, 1024 generation output limit, no SDK retries, frozen prompts/coverage rules, and default-off switches.
• Audit classification/generation stage and a bounded, cycle-safe exception type chain without exception text, arbitrary notes, credentials, or provider error bodies.
• Record response-header/body phase, known HTTP status before body reads, usage splits and finish reason in bounded synthetic acceptance transport. Do not record hidden reasoning.
• Tighten the acceptance request guard to require thinking disabled for generation too.

Verification

• Focused Hybrid tests: 49 passed.
• Complete offline suite: 632 passed in 48.14 seconds, with temporary database, dummy credentials and an external-network/private-data audit guard.
• Offline full acceptance rehearsal: all 8 cases PASS with MockTransport; exactly 14 embedding, 6 rerank, 6 classifier and 2 generation simulated attempts. This is wiring/guard coverage, not real model quality evidence.
• Ruff and git diff --check passed.

Scope and limitations

No paid calls, .env changes, public deployment, real ticket creation, timeout increase, prompt/rule tuning, or experiment-source changes. Historical failed acceptance report remains unchanged. Thinking was a configuration risk, not a proven sole cause of the previous timeout. Version/time semantic-audit warning and the 7 unexecuted real cases remain unresolved. Real retest requires fresh explicit authorization; missing usage must not be interpreted as zero billing.