Two open design problems in the work-stealing algorithm need detailed analysis: (1) no burst-handling data at 8+ carriers with load imbalance, and (2) wakeFirstIdle lowest-ID bias creating an ABA pattern where one carrier absorbs all steal work.

Issue 1: Sequential chain for multi-victim steal (docs/issues/multi-victim-sequential-chain.md)

• Explains Go's nmspinning approach: cap searchers at GOMAXPROCS/2, randomized victim start, steal-half semantics — all without per-P utilization measurement
• Proposes analogous options: cap at clusterSize / 2, rotating start index, adaptive cap based on idle count
• Key tension: with cluster size 4 and 2 carriers slow, the remaining 2 must both be available — soft limits at half-cores may be the ceiling

Issue 2: wakeFirstIdle lowest-ID bias and ABA (docs/issues/wakefirstidle-lowest-id-bias-aba.md)

• Long.numberOfTrailingZeros(word) always selects lowest-ID idle carrier → carrier 0 is perpetually hot-pathed while others stay cold
• Both bitmap (LIFO-like via scan order) and stack (LIFO by design) suffer the same ABA: carrier parks and immediately becomes next wake target
• Argues against dual structures (bitmap + stack) due to atomicity/consistency overhead
• Proposes single-structure solutions: rotating scan start, victim-ID-derived offset, bitmap+epoch for temporal ordering

// Current: always wakes lowest-ID idle carrier
int bit = Long.numberOfTrailingZeros(word);  // bias toward carrier 0

// Proposed (rotating scan):
int startWord = (scanCounter++ % words.length);  // distribute wake targets

These are issue drafts committed as docs — sandbox lacked permissions to create GitHub issues directly.