Repository navigation
feat(v4): enroll English emit/ingest round-trip in claim-witness corpus CI - #4947
Merged
Merged
Conversation
briansrls
changed the title
3 tasks done
briansrls
marked this pull request as ready for review
June 15, 2026 09:24
…us CI Rebased on main post-#4946: one additive corpus row on shard A; counts 45/19/26. Co-authored-by: Cursor <cursoragent@cursor.com>
briansrls
force-pushed
the
pr-4933-rebase
branch
from
June 15, 2026 10:07
e8daa7d to
13796d6
Compare
briansrls
added a commit
that referenced
this pull request
Jun 15, 2026
…9→20 (#4955) #4947 enrolled the English emit/ingest round-trip row into claim_witness_corpus_shard_a_rows (now 20 members) and bumped the full-list count datum, but did not bump claim_witness_corpus_shard_a_row_count. The shard-a gate's count guard (projected 20 != modeled 19) exits 2, so v4_claim_witness_corpus_a fails on every branch off post-#4947 main — a tree-wide merge block (confirmed: run 27548200843 shard-a FAILURE, shard-b success). Bump the datum to match reality (shard_a genuinely has 20 members). Minimal: only the consumed shard_a datum. The full-list count (54 vs 55 members) is an unconsumed/non-gating ledger and shard_b (35) matches, so both are left alone. Throwaway: the in-flight claim-witness-corpus model-first dissolution retires the count-guard datums entirely. Co-authored-by: Brian Searls <briansearls1@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This was referenced Jun 15, 2026
briansrls
added a commit
that referenced
this pull request
Jun 15, 2026
…ility) (#4959) * WIP: ci investigation * WIP: ci investigation * chore(ci): retire redundant *_row_count cross-checks (merge-skew liability) The claim-witness-corpus and affected-set node-frontier/testgen gates each cross-checked a projected row count against a hand-maintained *_row_count datum. That datum is a parallel ledger for a fact the typed shard List already carries: the gate projects rows directly from the list, so a dropped row simply is not enrolled and its witness is not run — the count datum cannot detect that and only DRIFTS under concurrent merges (a hand-count vs a hand-list is the merge-skew axis that reddened main on #4947: two PRs each list-consistent, the squash-merge re-resolve combined their list edits while the count stayed stale). Delete the cross-checks and the now-unused datums (CLAUDE.md ledger principle): - scripts/v4-claim-witness-corpus-gate.sh: drop count_data + expected_count fetch + the projected-vs-modeled comparison; keep row_count for the notice. - scripts/v4-affected-set-node-frontier-gate.sh: drop both expected_count fetches + both comparisons (node-frontier + testgen). - claim_witness_corpus_ci_runner.dag: delete claim_run_row_count (dead, 0 readers, and stale at "54" vs 55 real) + shard_a_row_count + shard_b_row_count. - affected_set_ci_runner.dag: delete ci_runner_node_frontier_claim_run_row_count. - affected_testgen_ci_runner.dag: delete affected_testgen_claim_run_row_count. Teeth retained: per-row pass + per-row perturb-RED + ExpectFail-drift + empty-roster fail-closed guards are all untouched. A real witness regression still trips RED. (Pass-by-omission on list-row deletion is restored properly by glob-discovery; this only removes the drift-prone parallel ledger.) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * WIP: ci investigation * WIP: ci investigation --------- Co-authored-by: Brian Searls <briansearls1@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
briansrls
pushed a commit
that referenced
this pull request
Sep 11, 2026
Keying the compatibility subject on the runtime's bytes rather than its enum variant was the right repair and it read a NonEmptyStr off the runtime row. For one of the three rows that string was never a digest: the DeepSeek image carried "aidendle94/sparkrun-vllm-ds4-gb10 (digest pinned in extdeps.docker.images.sparkrun_vllm_ds4_gb10)" -- a sentence about a pin standing where a pin belongs. A subject keyed on that compares a MUTABLE TAG as though it were an artifact, which is the same failure the byte-keying was introduced to remove, one level down. So the pin is its own question with two honest answers. PinnedByDigest IS the bytes; PinnedElsewhere names the authority that holds the pin and yields NO identity -- runtime_pin_as_identity returns an Optional and runtime_artifact_key renders a located <unpinned:...> marker, so an unpinned runtime cannot silently share a key with anything, including another unpinned runtime. The patch set is in that key, and it is the half the repair depends on. The next runtime this fleet needs is a derivative: upstream vLLM 385dce36 with the two vllm#55277 commits backported and a FlashInfer revision carrying flashinfer#4802 and #4947. Stock bytes and patched bytes can share a digest lineage, so a key that dropped the patch set would let the refusal measured on stock bytes refuse the patched image -- and the patched image is the entire repair. patch_set: [] on the three existing rows is a positive fact, which is why the official image's NoPE MLA refusal reads as a statement about upstream rather than about a local build. runtime_artifact_wire and runtime_artifact_key are deliberately two functions: a launch may legitimately pull an unpinned tag, and a receipt must never claim that pull was an identity. The new witness proves the discrimination without building anything -- same digest, same source revision, differing only in patch set, keys must not collide -- and its negative rows are live corpus rows rather than fixtures. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01W6hK2jVJ1neNMVR1oi6PyW
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Rebased english emit/ingest round-trip corpus enrollment on main post-#4940 (markup serializer keystone).
Adds
corpus_ci_gate_row_english_emit_ingest_round_tripto monolith roster + shard A; counts 43/19/24.Supersedes #4933 (stale pre-#4940 base).
Test plan
cargo test --workspace(local)