Skip to content

#127: the GGUF v3 and ONNX realizations in the kappa package (KA-07, KA-08) - #539

Merged
afflom merged 19 commits into
release/v0.5from
claude/issue-127-gguf
Nov 16, 2026
Merged

afflom merged 19 commits into
release/v0.5from
claude/issue-127-gguf

Conversation

@afflom

@afflom afflom commented Nov 14, 2026 •

Copy link
Copy Markdown
Member

Requested by Alex · project thread

Fixes #127 (L5.2). Per D7 (#110), kappa/ is a workspace source package outside language/ and schemas/, and this PR touches no language/ or schemas/ path. Its regenerated SPEC, model and record outputs land on release/v0.5 one at a time through the v0.5 queue.

Why this closes #127

#127 is accepted when every committed uor-addr fixture gets the same κ-label from LexLean. At the pinned uor-addr 0.2.0 there are 12 JSON fixtures (already done, KA-02), 3 GGUF fixtures (KA-07, this PR) and 2 ONNX fixtures (KA-08, this PR). With this PR the package gets the committed label on all 17.

Corpus and reference boundary (versioned)

  • Pin. Acceptance is bound to uor-addr 0.2.0 (crates.io, tag v0.2.0; GGUF and ONNX canonical form v2). The 17 fixtures are the ones committed at that pin: 12 JSON (tests/byte_identity.rs, KA-02), 3 GGUF and 2 ONNX (tests/fixtures/{gguf,onnx}, vendored byte-for-byte in tests/kappa/). The oracle pin does not move.
  • Python reference encoders. tools/canonical-gguf.py and tools/canonical-onnx.py at v0.2.0 print the committed label on all 5 binary fixtures, so at the pin the Rust crate, the Python encoders and the committed labels agree. LexLean tests against the Rust crate, which is the oracle linked in the build. The Python encoders are a second reference over the same 5 pairs and are not run in CI.
  • Pin vs current upstream (20a1c57, 7 commits past v0.2.0, not released).
    • The 5 binary fixture/label blobs are byte-identical at both refs.
    • Upstream adds a sixth pair, onnx/empty-name-tiebreak, under ONNX canonical form v3, which changes the Kahn tie-break for empty node names (d677d17). Its committed label sha256:984fe6a7… comes from v3; the v0.2.0 Python encoder gives sha256:7b6d2a01….
    • Upstream JSON now rejects duplicate member names (CT-B05).
    • Neither change is in a release. Both are out of scope for [0.5.0] κ-addressing in LexLean, validated against uor-addr #127 at 0.2.0, and adopting them is a later pin bump with its own R3 rows. [0.5.0] κ-addressing in LexLean, validated against uor-addr #127 comments refer to this as "the six-pair corpus versus the older five label strings".
  • Evidence. R3 rows KA-07/KA-08 (spec/src/IdsKA.lex.tex), scenarios (features/suites/kappa-addressing.feature), tests conformance_ka_07/conformance_ka_08 (evaluator + rustc differential + 17 planted defects). The product is the kappa package (kappa/src/Gguf.lex.tex, kappa/src/Onnx.lex.tex) with regenerated outputs adopted at 1ebca4ce (producer 43871301663 via regenerate/ka-539, from 3c037c9e, which merges release/v0.5 d07bd64c and relocks kappa). The earlier adoption at fd6b9e2c (producer 43781212846, base 7da6e937) is superseded. Earlier outputs from regenerate/ka07-ka08 and the full gate at ca8af2d4 are historical evidence for that head only, with their original scope.

R3 order (one commit each, per row)

  • KA-07, GGUF v3.
    • Register row (and KA-01 updated) → scenario → conformance_ka_07 → kappa/src/Gguf.lex.tex.
    • The test compares against uor_addr::gguf::{canonicalize,address*} on the 3 fixtures, inputs at every bound the crate checks, and seeded files. It runs under the calculus evaluator and as a rustc-built rust-std package.
    • 8 deliberately broken copies are each caught.
  • KA-08, ONNX (IR 1–13).
    • Register row → scenario → conformance_ka_08 → kappa/src/Onnx.lex.tex.
    • How the module works:
      • It reads the protobuf wire form only as far as the crate reads it (first-field scans versus full parses).
      • It writes graphs through an explicit task machine, so subgraphs need no recursion, at most 64 levels deep.
      • It orders nodes the crate's way: Kahn order, taking the last of the ready nodes that are least by name, op type and domain.
      • Other messages are digested over an explicit frame stack: fields in number order, and a payload that reads as a message is digested again, up to 32 levels.
    • The test compares against uor_addr::onnx::{canonicalize,address*} on the 2 fixtures, edge models (nesting depth 63–65, a cycle, IR 0/14, opsets, 10- and 11-byte varints, an unknown wire type) and seeded models, under the evaluator and under rustc.
    • 9 deliberately broken copies are each caught.
  • Oracle. The gguf and onnx features of uor-addr are turned on. The only lock change is one new edge, uor-addr → uor-prism-tensor (MIT, already in the graph).
  • Limits and spec text. kappa raises max_file_bytes to 16 MiB, because the rust-std certificate of onnxLabelOn alone is 5.3 MB. The §17.22 paragraph is updated and split with a cont item to stay within the 60-line item limit (NS-04).

Run locally: conformance_ka_07 passes (39 s) and conformance_ka_08 passes (225 s); all 5 fixture labels match. lexlean verify of Gguf passes with exact axioms.

Qualification

  • Current adopted head is 1ebca4ce. Its final-head gate is dispatch CI 43898718803. Bot-push runs 43899746321/43899746288 are action_required, not the gate. Superseded fd6b9e2c gates (43799050127 attempts 1–3, 43869262387) were failed or cancelled for runner, DNS, lane-attempt, atlas-rerun and base-move reasons recorded on the PR, and are kept separately.
  • Limitation: the R3 commits were not each observed red on the original history, and that history is not recreated.
  • Merge only on 1ebca4ce all-green, with the merge commit's tree equal to the gated tree. That closes [0.5.0] κ-addressing in LexLean, validated against uor-addr #127.

https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX


Generated by Claude Code

KA-01 names the module Gguf and its three roots.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Enables the gguf feature of the uor-addr oracle (adds only the edge to
uor-prism-tensor, MIT, already in the graph) and vendors the three
committed GGUF fixtures of uor-addr 0.2.0 beside their labels.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
The module Gguf writes uor-addr's canonical GGUF skeleton: header and
alignment, metadata sorted by key bytes with digested strings and arrays,
tensors sorted by name with relaid canonical offsets and digested data.
canonicalGguf, ggufLabel and ggufLabelOn are rust-std production roots.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
github-actions Bot and others added 5 commits November 14, 2026 22:05
KA-01 names the module Onnx and its three roots.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Enables the onnx feature of the uor-addr oracle (no lock change: the
edge to uor-prism-tensor is already there for gguf) and vendors the two
committed ONNX fixtures of uor-addr 0.2.0 beside their labels.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
The module Onnx reads the protobuf wire form as far as the crate reads
it, emits graphs through an explicit task machine (subgraphs without
recursion), orders nodes as the crate does, and digests other messages
over an explicit frame stack. max_file_bytes rises to 16 MiB for the
certificate of onnxLabelOn. The 17.22 paragraph is split into a cont
item to keep each item within 60 lines (NS-04); kappa sources are fmt'd.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
@afflom afflom changed the title #127 slice 4: the GGUF v3 realization in the kappa package (KA-07) #127: the GGUF v3 and ONNX realizations in the kappa package (KA-07, KA-08) Nov 14, 2026
@afflom afflom mentioned this pull request Nov 15, 2026
github-actions Bot and others added 2 commits November 15, 2026 00:43
The write recipes of SPEC.md §9.2, run by regenerate.yml (https://github.com/UOR-Foundation/lexlean/actions/runs/43647844292),
rewrote 396 paths.

The counts writer reported:

    BS-07 VECTORS: 410 -> 416 (crates/conformance/src/cases/bootstrap/rust_lower.rs)
    BS-13 VECTORS: 171 -> 177 (crates/conformance/src/cases/bootstrap/rust_loops.rs)
    BS-14 VECTORS: 414 -> 420 (crates/conformance/src/cases/bootstrap/rust_assemble.rs)
    BS-07: §31 row: 410 -> 416 (x1)
    BS-07: register: 410 -> 416 (x1)
    BS-07: scenario: 410 -> 416 (x1)
    BS-13: §31 row: 171 -> 177 (x1)
    BS-13: register: 171 -> 177 (x1)
    BS-13: scenario: 171 -> 177 (x1)
    BS-14: §31 row: 414 -> 420 (x1)
    BS-14: register: 414 -> 420 (x1)
    BS-14: scenario: 414 -> 420 (x1)

    VERIFICATION.md rows for the landing (counts | before | after):
    | `BS-07` VECTORS | 410 | 416 |
    | `BS-13` VECTORS | 171 | 177 |
    | `BS-14` VECTORS | 414 | 420 |

    For a reader:
      SPEC.md:8580 still states 410 (-> 416 for BS-07?) before "of `correspondCrate` (103 refused), and"
      SPEC.md:8752 still states 410 (-> 416 for BS-07?) before "per root: the committed target programs"
    wrote /home/runner/work/lexlean/lexlean/CONFORMANCE.md and /home/runner/work/lexlean/lexlean/ERRORS.md

    written, with SPEC.md and the model projected again; next: the parts without LEXLEAN_COUNTS
@afflom
afflom marked this pull request as ready for review November 15, 2026 06:21

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded acceptance readback at ca8af2d: this now includes both GGUF/KA-07 and ONNX/KA-08. The inspected cases compare real input bytes with committed labels and the pinned Rust oracle, exercise evaluator/canonical/axis/rustc comparisons, and require nonempty disagreements for the mutation controls. That is meaningful source credit; no algorithm defect was established in this bounded review, and no complete source/product clearance is claimed for the413-file change.

Run43672062503 succeeded only for plan, honesty and just-vv(rest):3 passed/15 skipped, including skipped Atlas/export/cross-host/Intel scope. Exact reproducibility passed. The existing full gate43685081548 is already queued for this head; use that qualification rather than starting another.

Before closing #127, make the versioned corpus/reference boundary explicit: the inspected tests bind17 fixtures to pinned0.2.0, while the issue also names the Python reference encoders and its prior discussion notes current-upstream versus pinned binary fixture differences. Link/reuse the existing reference and original R3/product custody that establishes the chosen scope; do not regenerate evidence merely to restate it. #127 remains open until that full-body disposition and resulting-head/actual-tree acceptance are complete. Existing implementation ownership remains.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The updated versioned corpus/reference section resolves the requested scope clarification. Independent upstream readback confirms the seven-commit difference, the same five existing binary fixture/label pairs, and the additional empty-name ONNX pair with a canonical-v3 compatibility change. This distinguishes pinned0.2.0 acceptance from a later pin bump without repeating fixture generation.

Retained independent reference captures establish the three GGUF Python correspondences. The two ONNX Python results are explicitly retained as owner-reported evidence; lacking a separate retained capture does not demonstrate failed or missing validation and is not a request to rerun them or invent a new receipt format. Link an existing correspondence record if available to make that evidence easier to reuse.

Prior meaningful test/plant/source credit remains scoped. The existing full gate43685081548 is still incomplete (65 successful jobs; Intel work remains), and this bounded body/corpus review does not supply a complete parser/product acceptance review or merged-tree equality. #127 remains open for its required final acceptance. No new check, producer or duplicate implementation was started. Receipt4771d347bc9a3fa0a980f6c9c1ec0d59e24084e39b3d8a67d2d4d5ab53ae57b6. COMMENT only.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 — the final-head verification record contradicts the implementation and closure scope. VERIFICATION.md:20029–20048 (blob7af8d4e) still describes KA01..06, says GGUF/ONNX forms are not written, and limits the binary fixtures to label-form checks. Please update this existing section to the actual KA07/KA08 scope and limitations, preserving/linking the genuine already-observed failures and mutation results. No new producer or recreated initial-red run is requested.

The independent remaining-source review found no concrete divergence in the inspected GGUF/ONNX algorithms against pinned0.2.0. Parent links and changed-file inventories confirm separate registration→scenario→test→implementation commits for both rows. Those commits establish source order; the reviewed VERIFICATION/body/commit messages did not independently locate the initial-red observations. This is not a claim that tests never ran or a demand for raw historical logs. Existing local green/plant claims and versioned-corpus/reference credit retain their stated scopes.

The smallest owner follow-up is the truthful documentation/evidence disposition above. Reuse the existing full gate and output custody; no duplicate reference capture or implementation. Full final-head/actual landing acceptance remains necessary before #127 closure. Receipt5c776c59d469b83af816ea88f76531c89df8fd5177259c6ff277b89bd5a28ae7. COMMENT only.

The section named KA-01..KA-06, said the GGUF and ONNX forms are not
written, and held the five binary fixtures to label form only. It now
states the KA-07/KA-08 scope (all 17 fixtures of uor-addr 0.2.0 by label),
the pin's limits, and the 17 plants with the counts of the passing local
runs at db30364 and f7671b1; no initial-red run was recorded.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Completed-gate credit at ca8: CI43685081548 attempt1 now has70 unique jobs and813 steps, all successful, including Intel. Exact-head reproducibility43672062358 also passed. These completed results should be reused; no fixture regeneration or gate replay is requested here.

The remaining next step stays with the source owner: correct VERIFICATION.md:20029–20048, which still says GGUF/ONNX are unwritten, and link/preserve the genuine original failure/plant observations as requested in review5910964523. Do not invent a red receipt or rerun a completed check to recreate history. Then disposition the final composition with current release/v0.5: the previously recorded synthetic7e75 was based on6c368, while the live release is now5c07. This gate does not by itself establish acceptance of a different merged tree. #127 remains open until the full delivered result is established.

Receipt pr539-ca8-completed-gate-20261115.json retains all job/step evidence. No execution, producer, retry, merge or CI control performed.

Generated outputs and locks taken from the release side; the regeneration rewrites them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

Pushed: 9c7db47 rewrites the VERIFICATION.md kappa section to KA-01..KA-08 (all 17 uor-addr 0.2.0 fixtures by label; pin limits: v3 sixth pair and duplicate-name refusal out of scope, Python encoders not run, 16 MiB max_file_bytes) and adds the 17 KA-07/KA-08 plant rows from the local runs at db30364/f7671b18. It states plainly that no initial-red R3 run was recorded; none is recreated. db48e4b merges current release/v0.5 (7da6e93, after #441/#545); generated outputs taken from the release side, and regeneration regenerate/ka-539 writes this head's outputs. The delivered head then gets one full gate.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded follow-up at db48e4b: credit 9c7db47's VERIFICATION repair covering KA01–08, all 17 pinned uoraddr 0.2.0 fixtures, canonical-v2 versus sixth-v3-pair/duplicate-key scope, and the 16 MiB certificate bound. The 17 KA07/08 rows retain owner-reported disagreement counts and implementation-commit observations; the missing initial-red R3 record is now stated explicitly. This is honest historical disposition, not recovered raw execution, proof that no execution occurred, or an R3 waiver.

Exact merge parents are 9c7db47 + current release/v0.5 7da6e93. Bounded blob equality preserves the previously reviewed Gguf/Onnx/kappa addressing/feature/IdsKA/manifest source, so no algorithm review or completed ca8 producer/gate was repeated. The compare response is capped at 300 files; this review makes no complete changed-path inventory claim. “Python encoders are not run” should retain the PR body's CI qualifier and the existing distinction between independently retained GGUF evidence and owner-reported ONNX evidence.

Release-side products/locks are provisional pending existing producer 43770177162 and its delivered descendant. At inspection it remained active with no artifacts; db48 CI was queued and repro active. Existing owner next step: bind delivered products to that producer, complete the resulting final-head gate, and establish actual merged-tree equality before acceptance. Prior ca8's successful 70-job/813-step gate remains credited only to its original input. #127 stays open.

Durable receipt: /workspace/work/pr539-db48-bounded-doc-integration-review-20261115.json; SHA256 dcd7a94b81a60dc3a239748c582b761b2df1447afbbc41e4809406a238fa8cb5. COMMENT only; no producer, retry, cleanup, or scope waiver authorized.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

New terminal diagnosis at unchanged db48: committed merge conflicts block delivery. kappa/lexlean.lock:3 contains <<<<<<< HEAD (blob47c8f78c); both producer oracles and records, plus repro, fail LLC0102 parsing that lock. kappa/expected/build/manifest.json:1 also contains <<<<<<< HEAD (blob05c74883); honesty fails parsing the committed generated manifest. These are concrete input defects, not an established memory or runner failure.

Producer43770177162 has identity/production-export success but oracles/records failure and combine/counts-stack/deliver skipped. It delivered no combined output. CI43770178630 and repro43770178529 fail; skipped platform legs and aggregate refusals are not independent passing qualification.

Smallest next step stays with the existing integration owner: resolve both committed conflicts against the actual composed KA configuration, inspect the remaining merge outputs for the same defect, then expose the corrected exact head and product custody for review. Do not hand-edit product identities or repeat the completed ca8 producer/gate. The previous bounded300-file comparison did not certify a clean full inventory. Documentation/source credit remains; final delivered products, final-head full acceptance and actual merged-tree equality are still owed. #127 remains open.

Immutable diagnosis /workspace/work/pr539-db48-terminal-diagnosis-20261116.json; SHA256 9ade7e48e454478039314d6546a12046752bac84654242b7cd7561b3aff0e955. Four new failed normative logs retained once; no root execution, retry, writer or CI-control action.

github-actions Bot and others added 2 commits November 16, 2026 00:16
…ock (release side; regeneration rewrites them)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3MAztfvN671cYrybev

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

Right, my merge db48e4b committed conflict markers in 20 kappa generated files and kappa/lexlean.lock (a batched checkout aborted partway). 3895d38 takes the release side for those 20 (all generated; regeneration rewrites them) and 7867e62 relocks kappa so project_config_sha256 matches this branch's kappa/lexlean.toml. No conflict markers remain anywhere in the tree (git grep over HEAD). Regeneration restarted on regenerate/ka-539 at 7867e62; producer 43770177162 is superseded.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The reported merge-conflict blocker is repaired at7867e625. Complete db48→7867 comparison has20files total:19 generated kappa/expected files restored exactly to accepted7da6 blobs, plus kappa/lexlean.lock. Authored KA source/tests/model/docs/workflows are unchanged. The current lock and generated manifest no longer contain the previously diagnosed conflict blocks; lock project_config_sha2562af6bcc6 matches this branch's configuration. Repro43781215224 passed. The owner's20-generated-files wording differs from the API inventory19+lock; this count discrepancy is not a new implementation barrier.

Current CI43781215250 fails honesty for the expected remaining reason: the provisional release-side generated record is stale against the repaired kappa lock/config (new retained log133024134749). Inventory aggregates refuse skipped execution; no independent platform failure or resource incident inferred.

Existing producer43781212846 has identity PASS and three writers queued at inspection. Preserve failed43770177162 separately; observe the existing producer, review its delivered input-bound products, then qualify the resulting final head and actual merged tree. No duplicate writer or completed ca8 gate repeat requested. #127 remains open; earlier R3 chronology limitation and Python-CI wording scope remain unchanged. Receipt /workspace/work/pr539-7867-repair-review-20261116.json; SHA256 e4209eafe104f7ceb6223f3cfc073bcb5e7410e60316985ba3304f9a8e827bf3. No full-repository marker scan or artifact ZIP equality claimed.

The write recipes of SPEC.md §9.2, run by regenerate.yml (https://github.com/UOR-Foundation/lexlean/actions/runs/43781212846),
rewrote 601 paths.

The counts writer reported:

    counts: no count moved

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Product adoption credited at fd6b9e2, the direct child of reviewed 7867e62. Producer 43781212846 completed all seven jobs, including counts/stack and delivery. Artifact 14195810780 is nonexpired and bound to the producer input. This is artifact metadata custody; no ZIP-byte equality is claimed.

The complete seven-page inventory has 601 unique changed paths, all generated kappa/expected products, including removals/renames. Authored sources, tests, model, documentation, workflows and lock are preserved. The manifest includes Gguf and Onnx and binds the current lock/config, resolving the inspected provisional input mismatch. Large omitted patches were not independently reviewed as bytes.

Current CI 43799050127 remains pending. Bot CI 43799065687 and repro 43799065556 report action_required, not success. Existing owner: update the body's old regenerate/ka07-ka08 and ca8 qualification references, complete final-head acceptance and actual merged-tree equality before closing #127. Preserve the honest original-red R3 limitation and prior reference evidence; no recreated history or blanket producer repeat. The prior ca8 full gate retains its original scope.

Receipt: /workspace/work/pr539-fd6b-product-adoption-review-20261116.json; SHA256 d1047851925250fa1c3b67f9334ad378536a6f2bbfc2a938d1fca627518f2d1b.

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

The aarch64 "build and non-Lean tests" job in gate 43799050127 failed with no test result. Its runner lost contact with the server during cargo test, and no log was written. This is a runner failure, not a code failure. When the rest of 43799050127 finishes, I'll rerun only that job in the same run. Merge still waits for every check on fd6b9e2c to be green and for merged-tree equality.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The ARM incident is recorded separately from the successful producer. Job 133079412399 in run 43799050127 attempt 1 is marked failed, while its cargo-test step has no terminal result. One log fetch returned BlobNotFound; check annotations were unavailable through the available access paths. Runner contact loss remains owner-reported. This evidence does not establish a code failure, but it also cannot prove that code was uninvolved; the underlying cause remains unconfirmed.

Preserve this failed attempt and the completed products. The existing owner retains the continuation decision; any resulting qualification must remain bound to fd6b9e2 and preserve actual merged-tree equality. No repeated producer or retry is requested or initiated by this review. #127 remains open.

Receipt: /workspace/work/pr539-fd6b-arm-incident-review-20261116.json; SHA256 49ab7c4170547610eca0a3ed3d650401ea55fc3802dbe763a02287bc6581fbb5.

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

The second failure in gate 43799050127 is "Intel conformance (print)", job 133093823238. All 20 tests in that lane passed (test result: ok. 20 passed; 0 failed, 6177 s). The upload of the intel-print artifact then failed with Failed to CreateArtifact: Unable to make request: ENOTFOUND, a DNS failure on the macOS runner. The tests didn't fail. The join needs that artifact, though, so the lane has to rerun.

When the run completes, I'll rerun its failed jobs once, this one and aarch64 job 133079412399, in the same run on fd6b9e2c. Merge still waits for all-green and merged-tree equality.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The new Intel print log independently confirms 20 tests passed, zero failed, in 6176.98 seconds; the subsequent artifact upload failed with CreateArtifact ENOTFOUND. Attempt 1 of 43799050127 finished with 67 successful jobs and three failures: ARM, print upload and Intel inventory aggregate. Preserve those distinct results: passing tests do not replace the missing artifact or qualify the refused join.

Attempt 2 now exists at unchanged fd6b and remains incomplete. Only partial attempt-2 job metadata was available, so no full count or cross-attempt all-green claim is made. The existing owner retains continuation and final-head/merged-tree qualification; no root retry, repeated producer or ARM-log refetch was performed. #127 remains open.

The PR body now correctly identifies the current delivered head/gate and the original R3 limitation, resolving the earlier stale-reference finding.

Receipt: /workspace/work/pr539-fd6b-print-attempts-20261116.json; SHA256 7973d85d64754591ca99eaa920f1569bf7e6e2e956207a97cc3f037699ba6803.

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

In attempt 2 of 43799050127, "Intel conformance (print)" (job 133153243376) refused to run: Intel host tests: foreign run, repository, attempt or construction job. The lane only runs against the construction job of its own attempt, and rerunning failed jobs doesn't rerun construction. That's the design working. It means a print lane can't be qualified by rerunning only the failed jobs.

I'll let attempt 2 finish so the aarch64 result is recorded. Then I'll rerun all jobs of 43799050127 (attempt 3) at the unchanged fd6b9e2c. That attempt is the qualifying gate. Attempts 1 and 2 stay on record as they are.


Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Attempt 2's new print log independently confirms that download succeeded, then the repository/run/attempt/construction guard refused execution before output creation. Upload consequently found no files. The precise mismatching artifact field was not independently decoded; the owner's attempt-mismatch explanation is consistent with the source checks and inherited construction timestamps.

The complete 70-record attempt-2 inventory currently has 67 successes, two failures and ARM still running. Those successes include inherited attempt-1 execution timestamps; they are not 67 new executions or a full-gate pass. Preserve attempt 1's passing print tests/upload failure and this guard refusal separately.

Existing owner retains continuation. Any subsequent qualifying attempt needs coherent same-attempt construction, lane receipts and inventory/join evidence, with final-head and actual-tree checks. Do not relabel old artifacts or weaken the guard. No root retry or repeated producer was initiated.

Receipt: /workspace/work/pr539-attempt2-inventory-guard-20261116.json; SHA256 c10138ede58c9f80efd1332ae97cfb19fa9bca61025d46fd1e37bd45cd4ebc03.

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

Attempt 2 finished. aarch64 job 133153243176 passed on fresh execution, and print and join failed as described above. Attempt 3 of 43799050127 is now a full rerun at fd6b9e2c: construction, every lane, inventory and join in one attempt. That attempt is the qualifying gate.


Generated by Claude Code

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

Attempt 3 of 43799050127 can't qualify the head. Atlas jobs census-d (133274946091) and cert4-replay (133303022254) failed with LLI9001 ... ci-parts/<shard>: the output directory of a part must be empty. Each atlas job downloads with pattern: atlas-*, and in a rerun attempt that pattern also matches the atlas-part-* artifacts left by earlier attempts. Each shard then finds its own earlier part already present in its output directory. The guard is working as designed, and nothing in this PR causes the failure. A same-run full rerun can't produce a coherent atlas.

I cancelled attempt 3 once this was established; attempts 1–3 stay on record as they are. A fresh gate on the unchanged fd6b9e2c is dispatch 43869262387. Its construction, lanes, atlas parts and join all come from one run and one attempt, and it is the qualifying gate. Merge only on that run all-green, with actual merged-tree equality. The rerun hazard is a CI issue, and I'll raise it with the CI owner.


Generated by Claude Code

github-actions Bot and others added 2 commits November 16, 2026 12:53
Generated kappa outputs and stdlib records take the release side for
regeneration; the register count is 524 (522 on the line plus KA-07/KA-08).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

#553 has landed on release/v0.5 as d07bd64, so fd6b9e2c would no longer merge to its gated tree. I cancelled dispatch 43869262387 on that superseded head.

  • 156f0fd merges d07bd64:
    • generated kappa/expected and the four conflicted stdlib records take the release side, ready for regeneration;
    • the register count is 524: 522 on the line plus KA-07/KA-08. validate-spec-links reports 524 rows bijective with the register.
  • 3c037c9 relocks kappa, because its lexlean.toml differs from the line.
  • Regeneration is now running from regenerate/ka-539. Then come a full final-head gate on the delivered head and actual merged-tree equality before merge.

Generated by Claude Code

The write recipes of SPEC.md §9.2, run by regenerate.yml (https://github.com/UOR-Foundation/lexlean/actions/runs/43871301663),
rewrote 388 paths.

The counts writer reported:

    counts: no count moved

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Product adoption credited at 1ebca4c, direct child of 3c037c9. Producer43871301663 completed seven jobs/76 steps successfully. Final artifact14239533258 is bound to that input (digest2c1e6da2…); this is metadata custody, not downloaded ZIP-byte equality.

The complete four-page inventory contains388 unique paths, all under kappa/expected, including old rename paths. Delivery preserves authored sources. Across the preceding integration with actual d07, inspected GGUF/ONNX sources, kappa conformance case and configuration are byte-identical; the subsequent relock changes only the project-config digest. The delivered manifest includes Gguf/Onnx and matches independently hashed current lock4ee639f1… and configuration2af6bcc6….

Existing owner: update the body from superseded fd6/canceled43869262387 to this delivery and current dispatch43898718803. Reuse that gate; bot CI/repro43899746321/43899746288 are action_required. Final-head acceptance and actual merged-tree equality remain pending. Preserve prior failed/canceled attempts separately and the honest original-red R3 limitation; no new producer or recreated history is requested. #127 remains open.

Receipt pr539-1eb-product-adoption-review-20261116.json; SHA2567217a849ce84c64b6a8e643b179a1536166e99ef1fb4970b8dcbbe8e21f2fa72.

@afflom
afflom merged commit 361dc35 into release/v0.5 Nov 16, 2026
70 checks passed

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

Merged as 361dc35 after the final-head gate on 1ebca4c (dispatch 43898718803) passed all 70 jobs. The merge commit's tree, 0079156b, equals the gated tree.


Generated by Claude Code

afflom pushed a commit that referenced this pull request Nov 16, 2026
…solation

Generated outputs take the release side (ours where the release has none)
for regeneration. SM-43 and SM-44 are both registered: 525 IDs. The
compiler-semantics ID is the line's until regeneration relocks it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Final technical landing credit: #539 merged as 361dc35 with ordered parents d07bd64 + 1ebca4c. Actual and final adopted head have the identical tree 0079156bb9ca95b2fb1958e5595fe60500d4414b. Dispatch 43898718803 attempt 1 checked out exact 1ebca4c and completed all 70 jobs and all 813 steps successfully, including honesty, Atlas/native acceptance, Windows, ARM Linux/Apple and Intel execution/inventory. Existing source, producer and product-adoption receipts are reused. Actual reproducibility 43947436552 passed its one job/seven steps; actual CI 43947436759 is planner-only (1 PASS, 17 SKIP), so its green status is not a second full gate.

The refreshed body correctly names the delivered head/current gate and pins acceptance to uor-addr 0.2.0's 17 committed fixtures. Owner comment6927965732 closed #127 after the merge; that external acceptance/closure is recorded. The body also explicitly preserves that the original R3 commits were not each observed red. Complete final green and ordered history do not replace that missing original observation; no waiver or historical reconstruction is inferred, and unqualified full R3 acceptance is not claimed. Existing Python evidence distinctions and all prior failed/canceled/successful receipts remain intact. Issue316 and old TC03/R25 STOP/future338/438/541 remain unchanged. COMMENT only.

afflom pushed a commit that referenced this pull request Nov 16, 2026
Vector counts add both lines' fixtures: rust_lower 419, rust_assemble 423.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants