fix: concurrentPool buffered-delivery stall and unbounded prefetch

[ppeeou]

Summary

Fixes three empirically reproduced defects in concurrentPool (regression tests committed first, all failing on the previous implementation):

• Buffered-delivery stall: next() only relied on a future fetch settling to trigger delivery, so once the pull chain drained, already-buffered results were never handed out while the following upstream fetch hung. next() now serves the buffer directly.
• Unbounded prefetch with a slow consumer: every next() launched length new fetches regardless of buffered results (upstream pulls == consumed × pool size). Prefetch is now capped at the pool size when no consumer is waiting; under active demand the pool stays saturated, which is what order-preserving pools (p-map included) must do past a slow head item.
• Post-error pulls (the long-standing TODO): the source is no longer pulled after it has thrown, and once nothing is in flight and the buffer is drained, later consumers receive done.

Throughput parity re-verified after the fix: p-map 455ms vs concurrentPool 455ms on the uneven-duration benchmark (concurrency 2) — the migration guide's claims still hold.

Test plan

• 3 new regression tests (hang via Promise.race timeout, produced ≤ consumed + pool, zero post-throw pulls), all red before / green after
• Full jest suite passes (140 suites, 1258 tests); lint 0; compile:check 0
• p-map parity probe on the built output: 455ms == 455ms

🤖 Generated with Claude Code