handover: mark the B2 items already measured, correct two assumptions about the workstation

[bernardladenthin]

docs/handover/local-agent-pr-a.md asked the local agent to discover seven things about the Windows
CPU variants. Four were already measured on that same workstation on 2026-10-09, so each item
now says MEASURED or OPEN with the numbers.

Verifying a MEASURED line is still worth the minute it costs — a different result there is a
finding — but the time belongs to the open ones.

Already measured, now stated

Item	Result
which clang	plain clang 23.1.3, all 14 variants, 0 errors
the directory	18 DLLs, nothing left in Release/, jllama-files.txt naming 15
verify-native-deps.py	18 libraries, 0 violations

Two assumptions that do not hold on this machine

Both would have cost the local agent time:

• the prompt said "LLVM 20.1.8 from PATH" — PATH carries 23.1.3, and Visual Studio ships
  no clang here at all (neither 2022\Community nor 18\BuildTools has VC\Tools\Llvm\x64). So
  the "newer versus older clang" choice it set up does not exist; the 20.1.8 of the earlier
  comparison was unpacked separately. 23.1.3 is the stricter one, and patches/0017 is what makes
  quants.c compile there.
• it asked "whether jllama-extras.txt exists" — it is absent locally by design. CMake never
  writes it; merge-native-artifacts.sh derives it in CI from everything jllama-files.txt does not
  name, which on Windows is exactly ggml.dll + ggml-base.dll (no $ORIGIN there, so they must be
  pre-loaded by full path). Asking neutrally invites a false defect report, so the item now says what
  the absence means.

Two additions the measurements make useful

The Java suite is marked as the item that decides whether the variant path works at all, with
the reason it is open: the measuring session's mvn compile dies in an Error Prone / NullAway crash
on args/package-info.java — unrelated to the native work, and disabling annotation processing then
breaks Lombok, so there was no quick bypass. CLAUDE.md's warning is quoted at it: a successful load
proves little, since in the earlier Windows measurement every crash came after a load that
reported the right device count.

The variant-selection step gets a caveat before the agent spends time on it: CLAUDE.md records
that ggml logs module decisions at GGML_LOG_DEBUG only, so -lv 4 may not name the chosen module;
listing the DLLs the JVM actually has mapped is the fallback. And the throughput item gets the
static-build figures (pp512 391.6, tg128 82.3 t/s, clang with GGML_OPENMP=OFF) as the line a
variant build should land near — so the number means something when it arrives.

Why this branch

Committed off main rather than onto claude/hopeful-pascal-9jlbqb: that branch carries an
unrelated open PR (#497, BoringSSL), and this file already lives on main, so the correction reaches
the session branch on its next merge without mixing concerns into someone else's review.

Verified

• check-shared-files.py clean (the file is not shared), buildcheck suite 113 green
• docs/handover/** is already covered in REUSE.toml, so no licence header is needed
• no control characters introduced; the protocol section is untouched

🤖 Generated with Claude Code

https://claude.ai/code/session_01E2h8gXyyE5UeimkL9vQv9G

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud