@@ -6,27 +6,42 @@ deliberately left alone. It is written for whoever merges the branch into
 session_01GBrvY1uPRbonUrDncT7rc7, 2026-09-29) could not reach openpocl's
 canonical copy.
 
-These are branch notes, not part of a release. When merging, move them to
-openpocl's docs (beside `docs/outline/40-traps.md`) or delete them, so
-`crmvec-publish.sh` doesn't publish them. Each later commit on this branch
-adds its entry here.
+These are branch notes, not part of a release. `crmvec-publish.sh`
+publishes only the files on its list, so this one stays out unless someone
+adds it; its content belongs in openpocl's docs (see "Where this belongs in
+openpocl" at the end). It is not called `DECISIONS.md` because openpocl's
+`docs/DECISIONS.md` is a generated file that sessions are told never to
+edit. Each later commit on this branch adds its entry here.
 
 ## Before merging
 
-1. **The branch was made on the public repo, not in `openpocl/harness/crmvec`.**
-   Copy its tree into `harness/crmvec` before the next
-   `crmvec-publish.sh`, or the next publish reverts it. Six files are new:
-   `crmvec-fpenv.c`, `crmvec-fpenv.h`, `crtest-ftz.h`, `crtest-own.h`,
-   `port/crmvec-port-e.c`, `port/pow-parity.h`. Then
-   `crmvec-publish.sh --check` should report a match.
-2. **It contains `main` as of `5bf6f82`** (merged in `357fa92`, no
-   conflicts) and was 0 commits behind it on 2026-09-29, so `main` can
-   fast-forward to it. `git log --oneline
-   main..origin/claude/beautiful-ramanujan-4gvzzd` lists what it adds.
-3. **The user asked that this branch not be merged by the cloud session**
+1. **Bring it in through the harness, not by merging the branch.** The
+   branch was made on the public repo, not in `openpocl/harness/crmvec`,
+   and public crmvec is written only by `crmvec-publish.sh`:
+   - copy the branch's files into `harness/crmvec` (or the next publish
+     reverts them);
+   - add the six new files to `crmvec-publish.sh`'s file list:
+     `crmvec-fpenv.c`, `crmvec-fpenv.h`, `crtest-ftz.h`, `crtest-own.h`,
+     `port/crmvec-port-e.c`, `port/pow-parity.h`. The Makefile needs all
+     six, so a publish without them builds nothing. This log is not on the
+     list, and needn't be;
+   - run the fresh-copy gate and the privacy scan as for any publish, then
+     publish; `crmvec-publish.sh --check` should then report a match.
+2. **Don't fast-forward `main` to this branch.** Its 11 commits are authored
+   `Claude <noreply@anthropic.com>` and carry a `Claude-Session` trailer.
+   Every commit on `main` is authored `anun333 <anun333@posteo.net>`, with
+   only the `Co-Authored-By` trailer, which is openpocl's rule for anything
+   public. The cloud session was set up with that identity and didn't know
+   the rule until it saw openpocl's CLAUDE.md. Publishing through the
+   harness gives `main` the usual author; the branch can be deleted after.
+3. **It contains `main` as of `5bf6f82`** (merged in `357fa92`, no
+   conflicts) and was 0 commits behind it on 2026-09-29. `git log
+   --oneline main..origin/claude/beautiful-ramanujan-4gvzzd` lists what it
+   adds.
+4. **The user asked that this branch not be merged by the cloud session**
    ("let the other maintainer version of you do that"). Nothing here was
    pushed to `main`.
-4. **CI** is green on every pushed head: `6bc7ff0`, `447d3c0`, `bed74e9`,
+5. **CI** is green on every pushed head: `6bc7ff0`, `447d3c0`, `bed74e9`,
    `41e9443` and `5fe2314` (run 36593612771).
    - GitHub's x86-64 runners vary: the `447d3c0` run had AVX-512, the later
      three didn't (their AVX-512 steps took 0 s).
@@ -302,6 +317,48 @@ exit non-zero.
   `ec1904b` on, a build of this branch links with conda's `LDFLAGS`. That
   was checked here with simulated flags, not by conda-build.
 
+## Where this belongs in openpocl
+
+Mapped from openpocl's CLAUDE.md and `docs/HANDOFF.md` as of 2026-09-28,
+the latest the cloud session saw. Run `harness/prior-work.sh` on each
+subject first (ftz, avx512, port_isint, variant_cc, multi-arch), in case
+the laptop has something newer.
+
+- **The crmvec record** (each commit above): `docs/own-math-library.md`.
+  The Cascade Lake timing goes in `docs/port-speed.md`. It bears on two
+  open lines there: making the x86 portable build the default (the owner's
+  call), and x86's 1.11-1.14x stragglers. On Cascade Lake the gcc build's
+  worst are `erff` 1.17x and `atan` 1.15x; the clang build's `sinhf` 1.13x
+  and `cbrtf` 1.11x.
+- **The CORE-MATH flush-to-zero findings:** `docs/pr-fix-intents.md`, the
+  CORE-MATH report entry, as a follow-up in the thread where CORE-MATH
+  answered the undefined-behaviour report on 2026-09-28. It is
+  outward-facing, so it needs the owner's yes. The draft and its `repro.c`
+  were written in the cloud session's scratchpad, which doesn't outlive the
+  container; both were sent to the owner through the app on 2026-09-29, to
+  go in `docs/drafts/`.
+- **Traps**, for `docs/outline/40-traps.md`, numbered after 82 (the
+  highest in the 2026-09-28 handoff):
+  - A check that finds its library through a RUNPATH of `$ORIGIN` tests
+    whatever `LD_LIBRARY_PATH` puts first, and prints that library's
+    verdict (`mpfrcheck`, `f16check`).
+  - `(a + 1.5*2^52) - 1.5*2^52` rounds exactly only for |a| < 2^51, so an
+    integer test built on it misjudges [2^51, 2^53).
+  - gcc's `-ffast-math` startup code leaves FTZ/DAZ (x86) or FZ (aarch64)
+    on for the whole program, and CORE-MATH is not correct under it.
+  - clang calls `_ZGVd*` for `-mavx` code. glibc's are IFUNCs, so a libmvec
+    replacement must check the CPU there.
+  - A check that ends in `| tail -1` or greps for a verdict passes when it
+    tested nothing (`crtest` and `mpfrcheck` typos, `hypotf-midpoints`
+    finding nothing). Say VOID and exit non-zero instead.
+  - The compiler marks functions that take vectors with the variant
+    calling-convention flag (`VARIANT_PCS`, `VARIANT_CC`), but not aliases
+    of them (riscv64 `fmin`).
+  - A `Multi-Arch: same` package can't carry a script that names the
+    host's library directory.
+- **Open items:** the "Deliberately not done" list above goes to
+  `pr-fix-intents.md`'s Backlog.
+
 ## Appendix: Cascade Lake, the AVX2 entry points per function
 
 `crtest time`, L1 column, fastest of 3 interleaved runs per build