arena: reuse certified empty spares without repeated sweep work

Author: Frityet

bench/arena-empty-reclaimed-2026-09-05/README.md

@@ -0,0 +1,45 @@
+# Empty reclaimed arena certificate and reuse evidence
+
+The final design and validation audit is
+`notes/arena-empty-reclaimed-reuse-2026-09-05.md`.
+
+`reuse-v2/` contains the final separate-atomic-head candidate, its matched
+runtime-input manifest, final build/fixture/TSan results, settled accounting
+control, layout probe, and all 20 primary performance processes. Every final
+process completed. `performance-summary.json` derives from the raw stdout;
+the original harness reports best-of-five inner rounds, while the surrounding
+script runs three fresh filtered pairs and seven fresh post-insertion pairs.
+
+Root-level performance files and `negative-v1-*` identify the rejected first
+revision. It was faster but retained 329→853 arenas across five closure rounds.
+Its deterministic generic-allocation and typed-pair controls each chose a fresh
+mapping despite two available real spares. The final revision keeps exactly four
+arenas across 32 cycles of each control and 320 across all five real closure
+rounds. The proposed interior-link/tail scheme was rejected before implementation
+because it would race concurrent GC publication and owner adoption.
+
+Both A/B revisions compare against the same immutable normal baseline,
+`/tmp/lj-gc-jit-combined-20260905-6cxpl6mp/normal`. It predates the final public
+MARK-scope repair and later scalar-read changes. Every tracked runtime input
+matches each candidate except arena C/header. Local paths and exact binary/source
+hashes are in the metadata; the temporary trees and binaries are not committed.
+
+All measurements keep GC enabled and JIT on, pin processes to CPU 30, alternate
+the order within fresh pairs, and preserve a 45-second process bound. CPU 30 has
+no SMT sibling and uses the performance governor; frequency is not fixed. Other
+agents ran functional work on CPUs 0–15. Own validation overlapped the diagnostic
+pair but had ended before the final primary pairs.
+
+`gcdiag.c` is a separate native frontend linked against each unchanged static
+archive. It reads allocator lists only from the main owner with no GC workers or
+MT execution, and snapshots outside timing. The settled harness adds an explicit
+collection after each closure round so retained physical capacity can be checked
+alongside live-byte and cycle counters. The final frontend counts both reclaimed
+heads. Command arrays, environment, exit status, wall/CPU time and partial output
+are retained in the JSON records. Copy evidence before rerunning the scripts,
+which write their output beside themselves.
+
+The initial standalone build's runtime-assertion-helper link error and a later
+new-fixture HugeTab initialization failure are retained with their corrections;
+neither is omitted from the audit. TSan demotes only GCC's known unsupported
+atomic-fence instrumentation warning. No host settings were changed.

bench/arena-empty-reclaimed-2026-09-05/after-insertion-01-after.stderr

@@ -0,0 +1,3 @@
+benchmark               total_s      ns/op
+tab_insert_newkey        0.3574    1786.75
+closures_upval           0.7799    1559.77

bench/arena-empty-reclaimed-2026-09-05/after-insertion-01-after.stdout

@@ -0,0 +1,3 @@
+benchmark               total_s      ns/op
+tab_insert_newkey        0.3670    1834.86
+closures_upval           1.9940    3987.91

bench/arena-empty-reclaimed-2026-09-05/after-insertion-01-before.stderr

@@ -0,0 +1,3 @@
+benchmark               total_s      ns/op
+tab_insert_newkey        0.3602    1801.20
+closures_upval           0.7796    1559.24

bench/arena-empty-reclaimed-2026-09-05/after-insertion-01-before.stdout

@@ -0,0 +1,3 @@
+benchmark               total_s      ns/op
+tab_insert_newkey        0.3672    1835.92
+closures_upval           2.0180    4036.03