[shinaoka]

What this changes

#339 narrowed two sentences that #337 had left on main. One of them ended with "its call site was not traced". It has been
traced since, in #340, so this replaces the gap with the measurement instead of leaving a stale disclaimer on main.

Measured (in #340, at e756f13 and at b989876, two runs each, release, counting allocator with backtraces at a size cap):

• The 468080 B is not the four-array helper's request. It is one Vec<usize> of 58510 elements, allocated three times per
  symbolic analysis (1404240 B in total) inside cholesky::supernodal::ghost_factorize_supernodal_symbolic.
• The 64000 B helper (4 · n · size_of::<I>()) only sizes a MemBuffer; it never appears as an allocation of its own.
• The StackReq composition is reproduced exactly: 64000 + 79928 + max(32000, 48000, 96000, 64000) = 239928, and the "before"
  figure is or(332712, 239928) = 332712.
• The largest single allocation is identical at both revisions (468080 B); what #337 removes is the COLAMD request from the or,
  which is why the simplicial largest drops from 332712 B to 239928 B and the supernodal one does not move.

Both files keep the entry on one physical line: a GitHub table cannot continue a cell across lines, which is how an earlier revision
of #339 split one entry into five rows.

Where the numbers come from (and what is not in this commit)

The probe is a throwaway crate that was never committed: two /tmp clones of the checkout, each driven by a path-dependency
probe that runs #322's 2000 x 2000, nnz = 7990 program with a counting global allocator (hooks alloc/alloc_zeroed/realloc,
a re-entrancy flag so that recording is invisible to the counters, and a backtrace captured only at or above a size cap and only
inside the measured window). The method, the toolchain, both revisions' full counter tables (two runs each) and the allocation
backtraces are in the #340 report comment: #340 (comment).

What this means for a reader of main:

• Nothing in this commit, or in the repository, can reproduce the numbers on its own. The entry cites the measurement rather than
  presenting it as this commit's result.
• No test in the repository asserts an allocation size. The two tests that touch the arm selection only check which arm was
  chosen, or that the numeric result agrees across thresholds: faer/tests/public_sparse_qr_symbolic.rs:32-50 picks between the
  arms on a 2 x 2 input, and faer/tests/sparse_lu_components.rs:36-55 runs AUTO / FORCE_SUPERNODAL / FORCE_SIMPLICIAL on a
  small banded matrix. Neither forces the supernodal arm at n = 2000, so neither corroborates these figures.
• To re-check independently: run the Sparse symbolic LU/QR with an explicit column ordering still allocate the COLAMD workspace #322 probe at e756f13 and b989876 in release mode with
  CARGO_PROFILE_RELEASE_DEBUG=1 CARGO_PROFILE_RELEASE_STRIP=none, forcing SupernodalThreshold::FORCE_SUPERNODAL for LU and QR,
  and record the largest single allocation plus its backtrace, twice per case.

Verification

$ git diff --stat
 CHANGELOG.md | 6 ++++--
 PATCHSET.md  | 2 +-

Docs only. The numbers are the ones posted in #340; I did not re-run them here, and the entry says where they come from rather than
presenting them as this commit's measurement.

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 0
UNVERIFIED: 1

Critical (must not merge)

none

Needs changes

none

Unverified claims

• CHANGELOG.md:20-23, PATCHSET.md:195 — Independent reproduction of the allocation measurements remains unverified. Issue Trace the 468080 B supernodal symbolic allocation reported in #322/#337 #340 contains before/after traces showing three separate 468080 B allocations and repeated counter results, but the original temporary counting-allocator probe is not attached; I cannot read that probe to audit its re-entrancy guard or counting window. To reproduce independently, run Sparse symbolic LU/QR with an explicit column ordering still allocate the COLAMD workspace #322’s 2000 × 2000, 7990-nonzero input with identity ordering at e756f13 and b989876, in release mode, forcing supernodal LU and QR and recording allocation backtraces twice per case. The repository test at faer/tests/sparse_symbolic_colamd_workspace.rs:75-82 forces simplicial analysis, so its success cannot independently establish these supernodal measurements.

Checked and fine

• Recomputed 58510 × 8 = 468080, 3 × 468080 = 1404240, and the workspace expression yielding 239928.
• Traced the three array allocations through cholesky.rs:2822,2894,2908 and csc/mod.rs:1536; PATCHSET.md:195 remains one physical table row.
• Tests and allocation probes were not rerun; the PR’s build and stable-Ubuntu test checks were still in progress when inspected, so full CI success was not established.

---

Reviewed head: 6ebe34f2310515bcef46f07e328a70238328d606 at 2026-11-09T08:07:00Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.