# Hard-broken records — the "king of the hill" decoders whose throughput stood
# unbeaten for so long that edging them by even a hair is the headline, not the
# margin. When a `zoomies-*` row beats one of these on decode throughput, the
# bench renderer stamps a 🏆 on our cell in the README/benchmarks tables.
#
# Keep this list TINY — only genuine long-reigning records belong here. Everyone
# else rides the normal "bigger number wins" rule with no trophy.
#
# Fields per entry (keyed by the rival's bench workload name):
# reigning_since = year the record was effectively set (for the badge text)
# nominal_mbs = the wall it held (best observed decode MB/s), for reference
# metric = which metric the record is on (default "decompress_mbs")
# note = badge caption
[lbzip2]
reigning_since = 2011
nominal_mbs = 777
metric = "decompress_mbs"
note = "the fastest bzip2 decode in existence for ~15 years — first ever beaten"
# ─────────────────────────────────────────────────────────────────────────────
# MEASURED BASELINE OF RECORD — Loki (AMD Ryzen 9 7900, 12c, 61 GiB), 2026-07-12.
# Durable, git-tracked copy of the byte-verified head-to-head so it SURVIVES a
# warehouse nuke. The bench warehouse is disposable; THIS + git are the truth.
#
# bz2 decode, deterministic 1024 MB corpus (-9, min-of-3), decode → /dev/null:
# znippy-bz2 (ours) ~753–770 MB/s @ ~11.7 cores
# output = full 1,073,741,824 B, SHA-256 IDENTICAL to `lbzip2 -dc` (correct).
# C lbzip2 ~701–763 MB/s @ ~11.7 cores
# → OURS WINS ~7% on a clean/settled box (standalone, direct /usr/bin/time).
# A settled-box harness run read 755 vs 759 (≈ tied); the afternoon
# "we're behind" figures were transient thermal contention, not the decoder.
# single-core (MLP interleaved): znippy-bz2_st ~130 MB/s vs lbzip2_st 71,
# bzip2 62 → ours ~1.8× per core.
#
# H1 FIXED (2026-07-14, measurement-only) — the harness's ~8% over-report of
# lbzip2 (759 vs 701 standalone) was a THERMAL/TURBO-BOOST ORDER bias, NOT a wall
# arithmetic bug: examples/nornir-bench.rs timed each decoder's iterations
# back-to-back, so in the full roster lbzip2 (running after the long single-core
# serial bzip2) was sampled in a fresh all-core-boost window while znippy-bz2 ran
# earlier in a worse one. Fix = sample all decoders ROUND-ROBIN (interleaved
# min-of-N) so each is measured across the SAME windows, AND base decompress_mbs
# on /usr/bin/time's OWN reported "Elapsed (wall clock) time" (the exact standalone
# basis), not the Rust Instant that also bills the wrapper fork/exec. NO decoder
# hot path touched (L2). Quiet-window interleaved re-measure of the SHIPPED
# binaries: znippy-bz2 min 1.389 s (~737 MB/s) vs lbzip2 min 1.525 s (~671 MB/s)
# = ours wins ~10% wall; output SHA-256 e4544f53… IDENTICAL over the full
# 1,073,741,824 B. Authoritative host-stamped warehouse run deferred to the Loki
# bencher (box was contended by parallel builds at fix time).
#
# RENAME (2026-07-12): zoomies-lbz2 → znippy-bz2 — INTENTIONAL, KEEP IT.
# Names the rival/format it competes with (the lbzip2 bzip2 format) so readers
# see "our alternative to lbzip2", not a cryptic "lbz2". Same package format.
#
# ✅ H3 — RESOLVED upstream (nornir 49b3cb6 "docs(render): H3 — word-bounded
# compare=/self= token match, not substring"). The mashup `compare=` matcher was
# SUBSTRING-based and rival names are substrings of our arm names
# (bzip2/lbzip2 ⊂ znippy-bz2; jar ⊂ znippy-jar; zip ⊂ znippy-zip), so a render
# MIS-MATCHED rival rows onto ours (the bzip2 row printed OUR number). nornir's
# `token_matches` (src/docs/sections.rs) now requires an EXACT arm id or a
# WORD-BOUNDED segment, and `compare=` tries exact-match first. `bzip2` inside
# `lbzip2` is letter-adjacent ⇒ rejected. Rendering znippy-zoomies is unblocked.
#
# ⚠️ H4 — OSM-KATANA SELF-COLLISION — PARTLY resolved, ONE hole left.
# osm-katana/README.md registers TWO adjacent bench blocks — Multi-core
# `compare=osm-katana` and Single-core `compare=osm-katana_st`. The exact-match-
# first rule added with H3 means each block now claims its own row whenever BOTH
# rows are present in the run, so the ordinary render is correct today.
# REMAINING HOLE: the word-bounded FALLBACK still accepts a token that is a
# PREFIX with trailing variant segments — `token_matches("osm-katana_st",
# "osm-katana")` is true (the `_` after the token is a boundary). So if the plain
# `osm-katana` row is ever absent or already claimed, the Multi-core block
# silently renders the SINGLE-CORE row. Same hole for `znippy-gz` vs
# `znippy-gz_st` — which is why examples/nornir-bench.rs still has to push the
# all-core row BEFORE its `_st` sibling.
# FIX (nornir/src/docs/sections.rs, `token_matches`): anchor the word-bounded
# fallback at the END of the arm id — a token may match as a boundary-delimited
# SUFFIX, never as a bare prefix/middle. Every sanctioned folding case is already
# a suffix (`skade_s3` ⊂ `data_pipe_skade_s3`, `skade` ⊂ `oxr_skade`, `lbzip2` ⊂
# `zoomies-lbzip2`), and all existing `token_matches` tests stay green; only the
# `<base>` ⊂ `<base>_<variant>` direction is closed. NOT applied yet: nornir had
# uncommitted work from another agent at the time (2026-07-22), and the standing
# rule is do-not-commit-another-agent's-tree. Exact patch is in that agent's
# report; apply it in nornir, not here — in-repo renaming would NOT help.