records: land the data from 5 of the 15 open PRs, correct 4 stale claims, leave #471 open

[lilith]

Triage of the 15 PRs open from 2026-08-06 to 2026-08-10, run the way #514's round was: nothing
merged as written, nothing closed whose data exists nowhere else, stale claims corrected in place
with a date rather than deleted. Full index: benchmarks/pr_triage_2026-08-13.meta.

Records only. No decoder code, no crates/ behaviour change, no feature flags. Every closed
PR's code stays on its branch at the head SHA recorded in the index. Nothing was re-measured —
another agent holds the box for timed work in a sibling repo — so every figure here is cited from a
committed record, from a CI log, or counted from the tree.

First: I was told three supersession claims that do not survive checking

The task brief asserted these from titles and campaign memory. Stated plainly rather than softened:

1. "perf(tracker): bounded spin then yield — the shipping build spun without bound #471 is one of the four waiting-policy arms perf(#455): c256x2048 t=8 — a finer shard block and four lock-waiting policies, both refuted, with the residual named #504 measured as NULL" — WRONG. perf(#455): c256x2048 t=8 — a finer shard block and four lock-waiting policies, both refuted, with the residual named #504 ran its
   four arms on c256x2048 at t=8, and its own not-measured list
   (benchmarks/c256_contention_2026-08-11.meta:128) ends with t=16. perf(tracker): bounded spin then yield — the shipping build spun without bound #471 ran at n=16 on an
   idle box across t=1..16 and read null at t=1/2/4/8 — agreeing with perf(#455): c256x2048 t=8 — a finer shard block and four lock-waiting policies, both refuted, with the residual named #504 — and 0.955 (8bpc)
   / 0.960 (10bpc) at t=16, with p90 and max moving by the same factor. t=16 is the only cell on
   this host where threads (16) exceed cores (12 logical). Not a duplicate. Left open.

2. "perf(tracker): shard picture planes by COLUMN — a w x h block is ONE borrow at every thread count (0.785x at 8bpc t=8, scaling 4.74x -> 6.13x) #469's column-sharding is the shape The &mut [u8] recon-kernel refactor, WRITTEN — sound, lands on #481's unsound ceiling at all 8 cells, and +1.6 MB instead of +96 MB #482's column-compact ReconBand landed" — IMPRECISE in a
   way that matters. perf(tracker): shard picture planes by COLUMN — a w x h block is ONE borrow at every thread count (0.785x at 8bpc t=8, scaling 4.74x -> 6.13x) #469's mechanism is a tracker change: RowMap, per-record c0/c1 column
   bands, cols_meet, COL_ANY, StridedRows. None of those symbols exist anywhere on main.
   The &mut [u8] recon-kernel refactor, WRITTEN — sound, lands on #481's unsound ceiling at all 8 cells, and +1.6 MB instead of +96 MB #482's src/owned_recon.rs is an owned buffer proved by borrowck with no record at all, one
   band per WORKER not per tile, removing recon's registrations rather than making them
   column-aware. The thing that actually generalises "one record for a rectangle" is perf(tracker): exact strided-rectangle borrow records — measured, default-off, and the cost model they correct #505/The code-placement lottery is removable (16-byte alignment, 1.37% -> 0.18% spread); the CDEF rectangle pays -2.0..-2.5% at t=8; both rectangles flipped ON #513,
   and it covers the filter chain — a different population. perf(tracker): shard picture planes by COLUMN — a w x h block is ONE borrow at every thread count (0.785x at 8bpc t=8, scaling 4.74x -> 6.13x) #469's goal was reached by two later
   changes, neither of which is its mechanism.

3. "fix(soundness): the rectangle guard hands out ROWS, not the hull — same one registration, no overlapping &mut #472 is what perf(tracker): exact strided-rectangle borrow records — measured, default-off, and the cost model they correct #505 implements and The code-placement lottery is removable (16-byte alignment, 1.37% -> 0.18% spread); the CDEF rectangle pays -2.0..-2.5% at t=8; both rectangles flipped ON #513 shipped by default" — CORRECT. Verified:
   DisjointMutRectGuard::row_mut(&mut self) at crates/rav1d-disjoint-mut/src/lib.rs:1615, and
   Cargo.toml declares __lf_rect / __rows_rect as retained no-ops because "the rectangle is
   now the DEFAULT path (2026-08-11)". fix(soundness): the rectangle guard hands out ROWS, not the hull — same one registration, no overlapping &mut #472's other half — the macos-15-intel
compact_read_per_row overlap it analysed and could not fix without an x86 box — was
   independently found, root-caused and fixed as the x86_64 t=8: deblock compact_read_per_row races a concurrent write (identical extents) — 4/768 vectors abort, aarch64 clean #494 t=8 race (X64_APPLICABILITY.md §G).
   fix(soundness): the rectangle guard hands out ROWS, not the hull — same one registration, no overlapping &mut #472 was right that the defect was on main and not an artefact of its own change.

What this lands

from	files	why it could not just be closed
#487	docs/SIZE_SWEEP{,_T8}.md + 9 data + 11 scripts/perf/*	main cites SIZE_SWEEP.md on 26 lines across 8 files and does not contain it
#488	6 benchmarks/ + scripts/perf/paired_ab.py	AGENT_BRIEF.md §6 cites lf_vbatch_iso_2026-08-10_v1.tsv by name
#471	4 benchmarks/park_not_spin_*	a t=16 result #504 never measured, and a 6x correction to a figure on main
#450	tile_threading_guard_2026-08-06.meta + 3 .tsv.zst	three metas on main cite it by name, two for its vector recipe

SIZE_SWEEP.md is byte-identical between pull/487/head and measure/size-sweep-intercept
(30a7294) — sha1 eaff7b70… — so this closes #514's deliberately-left-open pointer rather than
forking it. Excluded as code: examples/avif_to_ivf.rs, and a stray __pycache__/*.pyc that
30a7294 had already dropped.

Four stale claims corrected in place, none deleted

• crates/rav1d-disjoint-mut/src/tracker_shard.rs:582 + docs/C256_CONTENTION.md §5 — both say
  the earlier TinyLock nulls were taken "where contention is ~0.02% of registrations". That is the
  wrong counter: it is wide_probe::N_SLOW, a count of BorrowTracker::add_slow (the
  poisoned / wide-record / multi-block path), divided by the registration population and quoted as a
  contention rate. Traceable to an add_slow count at verify_compose2_2026-08-08.meta:278.
  Directly counted on the same vector: 0.061% of acquisitions / 0.122% of registrations — six
  times, against a different denominator. perf(#455): c256x2048 t=8 — a finer shard block and four lock-waiting policies, both refuted, with the residual named #504's 0.264% for c256x2048 t=8 is a different cell,
  counted directly, and is explicitly not disturbed; nor is §5's "at most 10.7% is waiting".
• docs/AGENT_BRIEF.md §6 — the waiting-policy row re-titled "CLOSED AT t <= 8" with its
  scope stated, plus a new row for oversubscription carrying perf(tracker): bounded spin then yield — the shipping build spun without bound #471's numbers and what would settle it.
• docs/CONTENT_CLASS.md — its "SIZE_SWEEP.md is NOT on main" header is now false. Its
  recalled count ("eleven references") was also wrong: counted on the pre-change file, 13
  occurrences on 12 lines, 10 of them in the body. Third recall-not-count slip the campaign has
  had to fix.
• benchmarks/tile_threading_guard_2026-08-06.meta §Result 3 — "a REJECTED attempt, recorded so
  it is not retried". It was retried five days later and it paid, because the retry obeyed the
  constraint that section ends on (leave the dense scan byte-identical). Annotated, not rewritten:
  its two designs are still correctly refuted.

Left open

#471 only. Records landed, code not. What settles it: the t=12 / t=24 / t=32 ladder (queued
on that branch, never got an idle window) and x86_64, where spin_loop() is pause and the
scheduler is not heterogeneous, so none of it transfers. Do not re-litigate at t <= 8 — that is now
null on two cells and two vectors.

Closed, each with a comment saying why and where its content lives

#450 #451 #468 #469 #472 #473 #474 #475 #480 #481 #485 #487 #488 — the SHA-verified table is in
benchmarks/pr_triage_2026-08-13.meta. #451's head 761366e8 is already an ancestor of main
(landed via #454, 490ed78). Two conclusions on closed branches are flagged there as refuted by
later rounds: #480's residual attribution (it is code placement, #506) and #485/#488's
count-is-cost framing (cost tracks distinct shard lines visited, #505).

Gates

cargo fmt clean on both touched packages. The only non-docs/-benchmarks/-scripts/ change is a
doc comment in tracker_shard.rs — which does mean this branch touches
crates/rav1d-disjoint-mut/** and so fires the 13 path-filtered disjoint-mut-ci.yml legs,
including both Miri legs, unlike the other records PRs in this family.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 5.51%. Comparing base (fcf07c5) to head (bb72946).

Additional details and impacted files

@@          Coverage Diff          @@
##            main    #518   +/-   ##
=====================================
  Coverage   5.51%   5.51%           
=====================================
  Files         85      85           
  Lines      36094   36094           
=====================================
  Hits        1990    1990           
  Misses     34104   34104           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.