[raylei50653]

Refs #547 (does not close it). Stacked republish: the chore PR opened next, against this branch.

Do not merge before owner review.

• Nothing here declares licence compliance.
• Every open item (L-1, L-2, L-3, L-4, M-1) stays OPEN.
• distribution.status stays local-only, and owner_confirmation stays null.
• The checker refuses any other combination while an item is OPEN.

What this does (owner decisions, 2026-10-08)

Item	Change	Still OPEN because
L-1 nvJitLink, cuFile, nvshmem	The wheel License.txt is kept. The version-matched official texts ship too: licenses/terms/cuda_eula_13.0.2.txt and nvshmem_sla_3.4.5.txt, each re-derived from its snapshot.	Which text governs a wheel-received object. Shipping the official text does not make it the governing one.
L-2 cuDNN ×5, TensorRT	Both texts ship. The TensorRT term is recorded (§12.2: one year, auto-renew, 90-day notice). The SLA §2(iii)(d) "by means of the internet" question is new. A per-clause-scoped downstream draft is in shipping/DOWNSTREAM_TERMS.draft.md. It is not shipped and not in force.	Which text governs; what "agreements with your Customers" requires. A README is not an agreement.
L-3 libgomp	COPYING3 and COPYING.RUNTIME come from the object's source package, plus SOURCE.txt. The correspondence is verified by comparison, not by debuglink (see below).	Corresponding Source for a patchelf-modified object; RLE eligibility; how long to keep the source available. The SRPM is not yet mirrored, because there is no public release.
L-4 NVIDIA common conditions	Per-condition audit in three layers. --library-path is not counted as proof.	"Only accessed by your application": inward provenance is solid, but nothing technically stops other programs (files are 0644/0755). "Material" is undefined. AGPL (M-1) may subject the SDK to an open-source licence.
M-1 model root (new)	Recorded as an open item: the YOLO26s engine has Ultralytics AGPL-3.0 lineage (ADR 023), and the training-data terms were not checked. #549 handles the separation and does not change any licence status.	Not audited. Needs a commercial licence, a legal opinion, or a different backbone.

libgomp evidence

• Same code. The shipped libgomp.so.1 (e28fb289…) and AlmaLinux libgomp-8.5.0-28.el8_10.alma.1.x86_64.rpm:/usr/lib64/libgomp.so.1.0.0 (e985bcbb…) share build ID 2244452c…. 24 code and data sections are byte-identical at the same addresses.
• What differs is only the patchelf change. .dynamic, .dynstr and .dynsym differ: there is one added DT_RPATH, and .dynsym changes only in section indices.
• Source package. The RPM header's SOURCERPM is gcc-8.5.0-28.el8_10.alma.1.src.rpm (c94dbbd2…).
• Package hashes authenticated. Both package sha256 values were checked through signed repomd.xml → primary.xml. The key fingerprint was not checked independently.
• No rebuild. The SRPM was not rebuilt.
• Re-checkable. license_audit.py check --gomp-rpm … --srpm … re-runs all of this.

Code

• license_audit.json moves to schema v2, adding supplied_texts, open_items, corresponding_source and downstream_terms.
• license_audit.py:
  • deterministic HTML article extraction;
  • RPM, cpio, tar.xz and ELF readers;
  • three new checks (supplied_texts, downstream_basis, corresponding_source);
  • also renders the downstream draft.
• The install cmake installs the supplied texts and checks their sha256. Bundle and package checks expect them and pin them.
• README.txt: the files under licenses/ are notices, not an agreement.
• Docs: resolved config §22 (design, contract, acceptance) and closeout §6.1 progress.
• No change to the model, tracker, runtime, entrypoint, operator library, third-party set, installer or any benchmark claim.

Tests and evidence

• Formal run results/547_licence/full_664208d1/ (head = contract + republish): PASS.
  • E1, tree: 64 files; only README.txt and THIRD_PARTY.md changed, plus the 7 texts. static 12/12.
  • E2, audit: 7/7 checks, complete.
  • E3, unsigned package: 7/7.
  • E4, clean-container install: passes.
  • Negative controls N1–N5 fail as required: a changed or removed text, a tampered install source, the wrong SRPM, and a public status while items are OPEN.
• scripts/pre_push.sh on the combined head: all green (5350 passed).
• Runtime identity: the implementation axis moves 221cc253 → 8e758b23 (+2 files, 2 changed). The stacked chore PR republishes it. CI on this branch alone is expected to fail the staleness tests until the chore PR is merged in.

Remaining owner items (#547)

• L-1..L-4 and M-1 conclusions.
• The release minisign key: commit only shipping/package/minisign.pub, in a separate PR.
• The publish decision. It also fixes where the libgomp SRPM is mirrored.

🤖 Generated with Claude Code

[raylei50653]

Response to the pre-merge review

Every finding is fixed in 16104f4e. The replay at cec6af39 passes (docs §22.7–§22.9). The cmake change needed another republish, which is in stacked PR #554.

#	Finding	Fix	Negative control
R-1 (P2)	.bss/.tbss size not compared	check_corresponding compares (sha256, size)	Growing .bss by 64 now fails with section .bss differs. The old code passed it.
R-2 (P2)	repo_file unconstrained	The checker and install_third_party.cmake require repo_file == "shipping/" + file. A file name of . or .. is refused.	../../outside/…, docs/… and licenses/terms/.. are rejected by coverage. N6: cmake -P exits 1 on an escaping path and writes nothing.
R-3 (P2)	All items CLOSED + no owner_confirmation passes	New check_public_gate: any status other than local-only needs owner_confirmation, every item CLOSED, the corresponding source published and the downstream terms adopted. A CLOSED item needs an owner_conclusion. tarball fails a package that is not local-only when checked without --pubkey. Closing items does not open distribution.	Unit tests: the public gate's three failures, the signature rule, and the local-only package that may be unsigned.
R-4 (P2)	664208d1 unresolvable; evidence only in results/	docs/reference/native_runtime_licence_evidence_547.json records the full commit and tree, the relation to the rebased head (only §22.4–§22.6 docs differ), the evaluation, 50 artifact hashes, the release hashes and how to re-fetch the inputs. #554 is not rebased, so the replay head cec6af39 stays reachable.	—
P3	#547 checklist missing M-1	Added to the #547 body (model root section).	—

Replay results/547_licence/review_cec6af39/:

• The installed tree is byte-identical to the formal run's; static 12/12.
• Audit 7/7.
• N6 fails as required, and the N7–N9 unit tests pass.

pre_push.sh on the simulated merge of #554 with this branch is all green (5362 passed).

Still the boundary: this is local acceptance on one machine, not an independent replay. Every open item stays OPEN, local-only, and owner_confirmation stays null. Merging #552 must not close #547 or move it to PUBLIC_DISTRIBUTION_READY. Do not merge before owner review: #554 into feat first.