ci: qualify post-v0.4.0 correctness candidate

[leehack]

Scope

Part of #76. Qualify immutable upstream 73ab7599b553c03f6f5d2db24a18ad76f2eb36a3 through existing read-only PR-triggered owner validation. Production submodule remains 5266f24da75dc449bd56cbed7addb9c8e4a6a73e (v0.4.0). No release policy, pins, runtime source, publication or dispatch changes.

• Replace redundant explicit v0.4.0 candidate lanes with post-v0.4.0, preserving pinned jobs/check names.
• Candidate Android ISA and emulated KleidiAI, Windows ARM64, Linux wrapper, Linux x64/arm64 packaging/dlopen, Windows x64 mtmd link coverage.
• Three additional jobs (Linux x64/arm64 and Windows x64 candidate counterparts), bounded to 30 minutes. Existing candidate jobs otherwise reuse their machinery.
• Exact SHA fetch/detach/readback before candidate builds, read-only contents permission and nonpersisted checkout credentials.
• Durable public grammar sampler check: 1999/2000 accepted; 2001/malformed rejected. This model-free character-grammar initialization exercises the actual built public function, not a helper mock; it does not claim token-level/model inference.

Validation

• Independent Astra blocking-only QA: PASS on all four files; independently reproduced positive candidate and negative published-v0.4.0 sampler outcomes.
• actionlint, Python compile, git diff check: PASS.
• Native release provenance verifier: PASS.
• Existing exact candidate local owner build: macOS Release CPU+Metal CTest 3/3 with assertions active, exports 36/36.
• Matching unmodified bridge local build and state/image/ASR four-mode/TTS two-mode checks PASS, recorded in Qualify post-v0.4.0 upstream grammar and GDN correctness fixes #76. These are local tests, not hosted publication attestations.
• First hosted run: 13/14 PASS, including both Windows ARM64 lanes; Android compiled then correctly rejected its previously unaudited source fingerprint.
• Maintainer explicitly approved adding only the independently audited exact candidate ggml/Kai pair while retaining v0.4.0 and all ELF/function/disassembly restrictions. The corrective commit implements that exact pair contract; no independent-component allowsets or cross-combinations are accepted.
• Independent Astra corrective QA PASS. Focused ISA tests 19/19; full Python suite 119/119. Real-tree added-caller/mutated-kernel and unknown/cross-combination negative coverage included. All source-delta evidence is in docs/post_v040_qualification.md.
• Replacement exact-head hosted matrix at 2d3bdce4cc64cf97bf9f4b0358c6f631adf8a66f: 14/14 PASS, including both Windows ARM64 lanes. Android inspected 256,191 instructions and verified 2,187 SVE/SME instructions confined to 18 exact audited functions. Run: https://github.com/leehack/llamadart-native/actions/runs/34167970481 . Provenance run: https://github.com/leehack/llamadart-native/actions/runs/34167970483 .
• Final committed-head independent Astra QA PASS; clean checkout, unchanged production submodule, zero unresolved review threads. PR remains draft; no mark-ready or merge performed.

Risks and boundaries

Current stable is still v0.4.0. This does not relabel later upstream source as a v0.4.0 wrapper rebuild. A containing stable release or explicit future development-channel approval is still needed for adoption. #76 remains open for complete qualification and adoption. CUDA and older-Android hardware are unavailable; emulated dispatch is not a physical-device result. Candidate PR lanes do not constitute the full publishing platform matrix or hosted exact-Web-artifact qualification. Runtime behavior/docs support claims are unchanged. No known PR-caused regression; review threads must be rechecked after hosted CI.