record(ENG-RECORD-CONFLICT-SURFACES): index #1543, the projection gate that never tests key uniqueness

[localai-bot]

.agents/issue-index.md did not track #1543. Verified with a positive control rather than a bare grep, because a null grep proves the terms wrong before it proves anything absent: the index carries 563 rows on origin/main, issues/1217 resolves to exactly one of them, and issues/1543 resolves to none, on origin/main and on the head of every open pull request. 18 of the 19 open heads carry the index and answer the control with 1 and #1543 with 0; the nineteenth, PR #361, predates the file, so git show exits 128 there and the row cannot be present. The issue was filed and never indexed, so the three places AGENTS.md requires to agree, the index, the owning row's spec, and the pull request body, agreed on nothing.

This change appends one row and changes nothing else.

What #1543 says, and why the checker is not touched here

scripts/check-public-doc-tables.py gates the projection documents for size and shape, the section and paragraph counts, the cell cap, the shrink-only oversized-cell ratchet, and never for key uniqueness. A public document is defined as the keyed current projection, one row per key carrying what is true now, and no gate enforces the "one" part. On PR #1081 docs/BENCHMARKS.md carried the key **NemotronH paged forward** (MODEL-NEMOTRON-H-ABI-A2P, #810) twice with contradictory content, one row reporting a GB10 read of 4/24 with the re-run pending and the other reporting the same gate passing 96/96, and the checker returned rc=0 on that file.

That is a checker-semantics change, which AGENTS.md routes to its own row, spec, red-before test or mutation, and green-after evidence. So this change indexes the issue and leaves scripts/check-public-doc-tables.py untouched, and #1543 stays open.

Why the row names an owning row rather than a spec

AGENTS.md gives an index row two admissible shapes: it names an owning row ID, or it names a spec that lists the issue under ## Owed. ENG-RECORD-CONFLICT-SURFACES is already the standing owner of public-document projection-gate semantics. #460, #495, #498, #507 and #1055 all sit there, and #1055's own row records in words that redesigning the doc budgets is that row's scope. That group is named at exactly this width on purpose: the five do not all sit in one script, since #460's check_links is defined in scripts/check-agent-record.py and #498's retired MAX_README_CHARS lives in scripts/check-readme-structure.py, so calling them check-public-doc-tables.py semantics would be false of two of them. The owner is also live rather than historical: the row is READY, not DONE, and its spec already names tests/scripts/test_check_public_doc_tables.py as a red-before suite, so its declared test surface is the checker #1543 indicts.

The ## Owed shape was considered and rejected on its meaning. ## Owed names the spec whose own work left the debt, and no spec here did. The defect surfaced during the review of PR #1081, whose separate repair collapses the duplicate rows into one without closing the blindness, because the next append reproduces it. Naming an owner is also not a claim that the scope is already written, so the row says in its own text that the fixing row still owes the spec and the red-before, and it carries the two cautions the issue records: the assertion has to be scoped per table, since audit-live-rows --check reports three IDs living in two matrices, BACKEND-CPU, BACKEND-CUDA-SM121 and BACKEND-VULKAN, so a naive global check reds main on day one; and the red-before is available rather than hypothetical, because PR #1081's own pre-repair head contains a real duplicate.

What the review changed

The fresh review reproduced the premise, the duplicate count, the ownership mutation and the gates, confirmed the owner choice on stronger grounds than the first draft gave, and returned one minor finding: two enumerations inside the appended cell were imprecise. The cell is append-only, so it can never be corrected once it lands, and both were corrected on the branch before it did.

The first is the script attribution above. The second claimed the only duplicate-key checks under scripts/ were three named files. scripts/check-agent-record.py holds two more, verified at this head: check_matrices refuses a duplicate ID over the matrix inventories, and check_issue_index refuses issue #N listed twice, which is the refusal this very row had to be checked against before it was appended. That file also reads docs/BENCHMARKS.md, through markdown_files. The substantive claim survives and is now stated as what it always was, that no checker asserts uniqueness on a projection document, and check_matrices is named as prior art, because it is the shape the fixing row should mirror and an immutable row implying there was none would cost someone that discovery later. The prior art is cited as path::symbol rather than path:line, because ENG-RECORD-ANCHOR-RATCHET measured that every stale anchor found in its 2026-08-13/14 campaign was still in range, so a line number in a row that can never be edited rots silently.

origin/main advanced by two commits during the review, so it is merged here with an authored, trailered message rather than by --no-edit, which produces the trailerless commit the trailer gate rejects. The range against the merge base remains a pure append, check-issue-index-append-only.py reports zero removed lines, and the pull request still shows one added line in one file.

Evidence

Ownership is gated, not decorative. check_issue_index counts the rows that name neither an owner nor a spec ## Owed entry and refuses any count above UNOWNED_HIGH_WATER, currently 33. Mutating this row's owner cell to a dash in the worktree takes the count to 34 and reds check-agent-record.py with #1543 named in the message, which also proves the checker read this worktree and not the shared checkout, a distinction that has silently inverted verdicts here before. The tree was restored byte-for-byte afterwards and the mutation was confirmed to have applied before the run.

Green after, re-run on the corrected and merged head: check-agent-record.py rc=0, check-issue-index-append-only.py rc=0, and scripts/agent-preflight.sh --fail-on-skip with zero skips. The two trailer gates were confirmed to have RUN rather than read as green: origin/main resolved to a commit, so BASE_UNRESOLVED does not apply; git merge-base --is-ancestor exited 0, so neither ANCESTRY_UNKNOWN nor the behind-base arm applies; and git rev-list --count exited 0 printing RANGE_COUNT=3, a numeric count above zero, so RANGE_UNKNOWN does not apply and commit-trailers and commit-style both took their run arm over the one commit under judgement.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]