PIZD · 5.0.0-rc2 → 5.0.0-rc3
Less encoding work.
Identical archive bytes.
Complete API costs on 30 public development excerpts and six correlated temporal prefixes. The same ten density wins remain, at 699,730 bytes versus 726,532 bytes ordinary.
These inspected inputs informed development. This shared-host study compares PIZD versions; the competitor scoreboard retains its own measured version. The refreshed standard panel still has 0/18 encoding-speed wins against its best reference, including a reusable Zstd control.
Complete encode costs
Values sum source-median milliseconds. Fresh-thread contexts include encoder initialization. Reused-thread contexts retain codec scratch between calls. File IO and thread creation are outside timers; model discovery, competition and production integrity are inside.
| Context | Sources | Templates | rc2 ms | rc3 ms | Time change |
|---|---|---|---|---|---|
| fresh-thread | development | OFF | 612.60 | 451.41 | -26.31% |
| fresh-thread | development | ON | 813.64 | 524.09 | -35.59% |
| fresh-thread | temporal | OFF | 695.93 | 659.52 | -5.23% |
| fresh-thread | temporal | ON | 1030.13 | 653.22 | -36.59% |
| reused-thread | development | OFF | 612.60 | 390.99 | -36.18% |
| reused-thread | development | ON | 788.35 | 445.63 | -43.47% |
| reused-thread | temporal | OFF | 682.45 | 626.33 | -8.22% |
| reused-thread | temporal | ON | 992.91 | 634.86 | -36.06% |
Slower cases remain visible
The following medians increased in the main capture. Focused follow-up did not reproduce the three development regressions; both captures are retained. No universal latency win is claimed.
| Source | Context | Templates | Time change |
|---|---|---|---|
| 05_loghub_openstack.log | fresh-thread | OFF | +3.06% |
| 12_loghub_android.log | fresh-thread | OFF | +31.17% |
| 29_github_events_20150101_12.jsonl | fresh-thread | ON | +24.50% |
| wikimedia_pageviews.test.5242880.log | fresh-thread | OFF | +1.53% |
What changed
Reuse independent Zstd streaming workspaces, prepare template catalogs and column layouts once, and skip extra template trials when both bounded samples show high structural diversity. The two samples cover up to 32 KiB each. Repetition elsewhere can be missed; this is a routing heuristic, not an incompressibility proof. Workspace remains allocated until thread exit.
Verification and complete evidence
All 144 corresponding archive/context identities match before/after. The two accepted replays perform 2,016 exact/SHA restores including warmups. All 18 enterprise CLI controls retain identical archives and restore exactly. The full library/integration suite passes 273 tests across 57 targets.