[leehack]

Moves third_party/llama.cpp from upstream v0.8.0 (8642541b) to stable v0.9.0 (cd959bae, 2026-11-04, v0.8.0 + 150 commits) and audits its CPU ISA source pair.

Head ef32565, based on main f3961b7 (merged in, no rebase).

Merging this publishes nothing. Since #178, auto_native_release.yml runs only on manual dispatch. The 2026-11-05 automatic dispatch failed only at the Linux arm64 and Android arm64 CPU ISA audits (Unaudited ggml/src + kai source fingerprint pair ('fead7601…', 'fd5b9ca9…')) and published nothing. A maintainer dispatches v0.9.0 after this lands. No tag, release or dispatch was made from this branch.

Changes

• Submodule: v0.8.0 → v0.9.0 (separate commit).
• validate_wrapper.yml: "v0.9.0": "cd959bae…" added to LLAMADART_UPSTREAMS. It is the pin row, so it builds the checkout as is and runs the pinned-only steps; tools/ci_scope.py resolves upstreams = [v0.7.0, v0.8.0, v0.9.0], pinned = v0.9.0. The v0.7.0 and v0.8.0 rows stay (maintainer decision).
• Android/Linux arm64 ISA: the v0.9.0 ggml/src digest bound to the unchanged KleidiAI v1.32.0 kai digest, granting exactly the v0.8.0 function set.
• Wrapper: no source change. mtmd_gen_audio_info.n_voices and the removed mtmd_helper_gen_audio_inp.voice are not read; Chatterbox reports UNKNOWN/unsupported. Chatterbox and predefined voices are out of scope.
• Docs: docs/v090_qualification.md (wrapper, per-patch review, consumer ABI delta), docs/v090_android_isa_qualification.md (source pair, FHM re-check with 0005/0006), AGENTS.md pointers. docs/native174_f16_dot_accumulation.md no longer says a daily detector publishes v0.9.0.

Merge of main (#177, #178)

File	Resolution
.github/workflows/validate_wrapper.yml	Conflict. Took main's LLAMADART_UPSTREAMS form; added the v0.9.0 row. The old per-job v0.9.0/pinned matrix edits are gone with that form.
tests/test_ci_scope.py	Conflict. Took main's version; it already checks that the pin is exactly one row and that the v0.7.0/v0.8.0 rows remain.
tools/validate_android_cpu_isa.py, tests/test_android_cpu_isa.py, AGENTS.md, docs/platform_backend_strategy.md	Auto-merged: #177's +fp16fml decoding and test sit next to this PR's v0.9.0 pair.

Carried patches

Reviewed 2026-11-07 against v0.9.0 and upstream master f7591a6 (b12190). apply --strict succeeds on both and applies 0001, 0002, 0005 and 0006, the four that manifest --upstream lists. Every patch passes git apply --check on both. series.json is unchanged.

Patch	v0.9.0	Upstream	Status
0001 Adreno 750 matvec	Applied	#29165 open draft; use_subgroups unchanged on master	Keep
0002 buffer device address	Applied	No fix on master; still no entry-point check	Keep
0003 embedding-mode reserve	Not applied (upstream_fix, #31086)	Fixed since v0.8.0	v0.7.0 row only
0004 Metal unary cast	Not applied (upstream_fix, required_text absent, #31360)	Fixed after v0.8.0	v0.8.0 row only
0005 ARM F16 dot	Applied	#31734 open; NEON still accumulates in float16x8_t	Keep
0006 ARM F16 GEMM	Applied	Same; llamafile still tinyBLAS<8, float16x8_t, …>	Keep

Measured: v0.7.0 manifest lists 0001–0006 but strict apply rejects 0004 (required_text absent), so it builds 0001–0003, 0005, 0006; v0.8.0 lists 0001, 0002, 0004–0006. Deleting 0003/0004 means dropping the v0.7.0/v0.8.0 rows — a maintainer support-policy call.

ISA audit

Source review unchanged from v0.8.0 except three ggml SVE256 paths in arch/arm/repack.cpp, behind __ARM_FEATURE_SVE and a runtime 256-bit check, so not in the non-SVE variants. The source pair still fingerprints unpatched upstream (fead7601… / fd5b9ca9… recomputed on the merged tree), so 0005/0006 do not change it.

Re-check on the merged tree (0005/0006 applied):

Module	Audit	FMLAL/FMLAL2
Android armv8.2_2 (NDK 28.2)	PASS, 266,016 instr; 3,130 SVE/SME in the 22 functions	0
Android armv8.6_1 (NDK 28.2)	PASS, 269,285 instr; 3,130 SVE/SME in the 22 functions	10, all in ggml_vec_dot_f16
Linux baseline / linux_armv8.2_2 / linux_armv8.6_1 (ubuntu:24.04 arm64, GCC 13.3, check_linux_arm64_cpu_variants.sh)	PASS all three; QEMU selection correct on cortex-a53, cortex-a72, neoverse-n1, neoverse-v1, neoverse-n2; finite F16 dot check passes on all 5	0 / 0 / 10 (fhm=10/0)

validate_linux_arm64_cpu_isa.py allows fhm only with FP16 on ARMv8.4+, i.e. linux_armv8.6_1, not armv8.2_2. validate_android_cpu_isa.py only enforces SVE/SME containment (decodes FHM via +fp16fml) and CI/release run it on android_armv8.2_2 only; applying the Linux family policy to both Android modules also passes, with fhm allowed only for armv8.6_1.

Local evidence (merged tree)

Check	Result
ci_scope.py with the workflow table	upstreams [v0.7.0, v0.8.0, v0.9.0], pinned v0.9.0
macOS arm64 macos-arm64-full + tests (M4 Max)	ctest 85/85 incl. llamadart_f16_dot_accumulation_test; validate_exports.py --forbid-import __cxa_atexit 92 exports; check-apple pass
Python test_*.py / *_test.py / verify_release_provenance.py	446 OK (1 skip) / 23 OK / pass
actionlint	clean except the known windows-11-vs2026-arm label

Hosted evidence (head ef32565)

All 52 checks pass. Rows are now v0.7.0, v0.8.0 and v0.9.0 (#178 dropped the older rows). The v0.9.0 rows skip "Select candidate upstream" and run the pinned-only steps (Android bundle exception runtime, Linux static-destructor check), confirming it is the pin row. Includes the v0.9.0 Android ISA, emulated KleidiAI dispatch, Windows arm64 KleidiAI/Vulkan, Linux arm64 artifact contract (variant audit + QEMU), wrapper contract, macOS exit teardown (Release with the F16 conversion-path test, and ASan), and the Metal OS floor at the pin.

• Validate Wrapper: success
• Metal OS Floor: success
• Validate Linux Release Rows: success
• Validate Release Provenance: success

Not covered

• No lane runs the new ggml SVE256 repack paths (they need an SVE variant on 256-bit SVE).
• Real-model TTS smoke, and model, family and device qualification, are left to the consumer release.

Follow-up (not in this PR)

• tools/validate_android_cpu_isa.py still allows the v0.4.0–v0.6.0 source pairs although no LLAMADART_UPSTREAMS row builds them any more.

Before merge

• The 0.18.0 hold is cleared: llamadart v0.18.0 was published 2026-11-07 (leehack/llamadart#1746 merged).
• Independent adversarial audit: ACCEPT at exact head ef32565 on main f3961b7f, no blocking findings: submodule bump and merge resolution, per-patch table, v0.9.0 source pair recomputed (fail-closed; mutations of the pair, function set and upstream row all fail tests), Android ISA rebuilt (FMLAL only in armv8.6_1), exports 92, check-apple, CTest 85/85. Wording nits (AGENTS.md calls 0004 "retired"; patches README line 73 mentions the automatic release) go into the v0.4.0–v0.6.0 source-pair prune follow-up.
• Fresh CI if the head moves.
• After merge: manually dispatch the stable release for v0.9.0, check its assets, then land #176, then the llamadart pin and bindings.