Avoid ArrayList<Integer> boxing when collecting ParNew promotion-failure sizes

[java-dependency-upgrade-fixer]

What

GenerationalHeapParser.parNewPromotionFailedDetails collected per-thread promotion-failure sizes into an ArrayList (one autoboxed Integer allocation per matched '(N: promotion failure size = X)' block, plus ArrayList's own internal Object[] growth) and then copied/unboxed every element into a final int[]. This companion replaces the ArrayList with a growable primitive int[] (capacity doubling from 4) that is trimmed once via Arrays.copyOf, matching the exact same values, order, and empty-array behavior when the regex finds no promotion-failure-size block (the null-guard already present in the PR head is preserved).

Why

The PR mirrors microsoft#579 and, per the evaluation context, one changed parser path still gathered primitive promotion-failure sizes through ArrayList before allocating/copying an int[]. This is exactly that path (parNewPromotionFailedDetails, lines changed by the same PR to add the block == null guard). No other changed line in this PR (CMSPatterns.java's cached SPLIT_REMARK_REFERENCE_BUG_DURATION regex, or the other null-guard additions in GenerationalHeapParser.java) presented a similarly safe, material allocation/parsing-work reduction, so those were left unchanged.

Behavior preservation

• The private field promotionFailureSizesForwardReference is set to the identical int[] content, order, and length (0 for no matches, N for N matches) as before.
• No downstream code reads this field today (confirmed by repository-wide search), so there is no other observable surface; a focused reflection-based characterization test locks its value directly for 0, 1, and 28 (repository-fixture-derived) promotion-failure-size blocks.
• All other statements in the method (clock, GC cause, garbage collection type, memory pool summary, duration parsing) are untouched.
• ArrayList remains imported and used elsewhere in the file (an unrelated event queue field), so it was not removed.

Evidence

• Focused behavior command passed identically before and after the production edit (3/3 tests).
• Full parser module test suite (124 tests) passed identically before and after.
• Full repository reactor build (7 modules) passed after the change.
• Isolated microbenchmark of the exact collection-building logic (28-block case) shows a deterministic drop from 28 boxed allocations + ArrayList growth to 0 boxed allocations + 3 primitive array (re)allocations, with a measured ~27% reduction in median wall time (733.82 ms -> 537.30 ms per 5,000,000 iterations, 5 trials).

Limitations

• The end-to-end per-line parsing speedup is smaller than the isolated microbenchmark suggests, since regex matching dominates total cost and real GC log promotion-failure records typically contain only a few blocks.
• promotionFailureSizesForwardReference has no current downstream reader, so this is strictly an internal allocation-efficiency improvement, not a change in externally observable output.

---

Generated by JAIPilot Cloud for #1 from Anthropic session sesn_01UZHhdPrsSrQEbHSEmJDv9c.

[skrcode]

Campaign acceptance microsoft#5 (independently reviewed)

• Identity: companion commit c49248f has exact input head 2b427ea as its sole parent.
• Scope: seven production-line changes in the already-changed parser path plus three directly relevant characterization tests.
• Behavior proof: exact empty, single, and 28-value output arrays and ordering are pinned; 3/3 focused tests and all 124 parser tests pass before/after.
• Allocation proof: representative 28-block path moves from 28 boxed Integers + ArrayList growth to zero boxing + three primitive-array reallocations; five-trial isolated median 733.82 ms -> 537.30 ms per 5,000,000 iterations (~27%).
• Final verification: full seven-module repository reactor passed.
• Cost: Anthropic list $2.37; estimated total $2.40.
• Independent fork CI did not attach to the generated feature-branch draft; repository-native verification above came from the managed exact-head session.

Accepted as a genuine, reviewable internal allocation improvement; remains draft and unmerged.