Forked workers returning results over pipes can silently serialise

A fork-per-worker pool (P workers, each writing its per-task results to its own pipe; the parent waits on worker 0's pipe to EOF, then worker 1, and so on) only reached ~1.3 of 4 cores. Each task's result blob was close to the 64 KB Linux pipe buffer, so workers 1..P-1 blocked in write() while the parent was still draining worker 0. CPU-s over wall shows it before any profiler: CPU-s / wall well below P with no I/O wait. Fix: give each worker an unlinked tmpfile() (inherited across fork), have it write there and _exit, and have the parent waitpid, then lseek to 0 and read. Alternatives are poll() over all pipes, or results small enough to fit the buffer. Check with a fixture whose per-task output is larger than 64 KB; small test fixtures hide it.