[cppbear]

gen --tasks previously bounded only the result queue, while fix --tasks was not passed to its scheduler. Both commands now default to four active focal-function jobs and reject zero before loading model configuration.

• Generation holds each slot through result delivery; repair holds it through result saving. Candidate validation and repair source/target mutations stay serial while model requests can overlap.
• Import compilation for both direct and prefix generation runs in the serial consumer, and filtered imports are saved for later integration-test construction.
• Every worker is joined before cleanup. Failures skip statistics; repair restores sources, retains recovery material on failure, refuses existing backups, and deletes only its own successful backups. Temporary compiler inputs are restored on normal completion or unwinding.
• Integration coverage reports are removed after every function has read them. The new multi-function integration regression exposed the previous early deletion.

Validation: workspace build and tests passed locally (30 passed, one live-model test ignored), as did the minimal full pipeline and all 11 coverage compatibility cases. Real CLI measurements for N=1,2,4 reached the requested model concurrency and kept compilation serial. Scheduler/failure checks passed locally. The full Linux regression passed on both the branch run and PR run, including consumer-failure queue draining and preservation of unrelated source backups.

The new CI fixture uses a loopback model and real Cargo to cover queue backpressure, both integration generation paths, worker/validation failures, and source/backup preservation. No real model service is called. This change does not alter request retries or timeouts, add cross-process locking, or rewrite Git history.