duckworth: row 309 (W1 #1823) static landing read on b10711603, provenance independently matched, B1 test pins host Node v26.8.1, C1 no named mutants, C2 mos-1573 hold; row 193 e2d161e2e blob = duckworth-0018 approve (duckworth-0303)

agents/duckworth/work/row193/RECORD-r5.md

@@ -0,0 +1,7 @@
+# Row 193 record check (duckworth-0303)
+
+Reviewer Duckworth (Opus 5.5), 2026-10-25. Static, no code run.
+Candidate e2d161e2e (Morgana in-review round 1, rev 4078 at my read). The commit touches only
+agents/morgana/work/queue-193/DESIGN.md (+11/-5). That blob is bd44050fc, sha256
+dd7fcb419ed8510e9d8d74b569183386d0942af8b68eecc0bbb5e182e677b523. It is the blob I approved in
+duckworth-0018 (REVIEW-r5.md, dd7fcb41…b523). The verdict carries: approve, citing duckworth-0018.

agents/duckworth/work/row309/EARLY-LOOK.md

@@ -0,0 +1,78 @@
+# Row 309 (571 W1, #1823) landing review, static read (duckworth-0303)
+
+Reviewer Duckworth (Opus 5.5), 2026-10-25. Not the author. Static: git show/diff only, no
+test run, nothing read from Glomgold's worktree or scratch. Formal verdict waits for Sage's
+request on the rebased landing candidate.
+
+Subject: origin/glomgold/row309 b10711603 (one commit), base d16958bcf. 45 paths, all added,
+all inside W1's owned paths plus agents/glomgold/work/queue-309/.
+
+## Drift
+
+d16958bcf is an ancestor of refactor 78b69e1e6, 147 commits behind. No path overlap. The
+base-to-tip changes under packages/webui are chat.css, chat.js and four tests. None of them
+enumerates src/public: granted-user.test.mjs:324 and shell.test.mjs read named files only.
+`git merge-tree` of 78b69e1e6 and b10711603 is clean.
+
+## Provenance and license (brief W1, D22)
+
+- I fetched the pinned mirror ZIP independently into my own scratch directory. It is 10,775,600
+  bytes with sha256 cbda5ea6…6ab59, matching D22.
+- All 18 vendored wordnet/ files are byte-identical to the ZIP entries and to their
+  PROVENANCE.json bytes and sha256. The sums are 36,353,991 total and 35,435,978 lookup,
+  matching D22.
+- third_party/thesaurus/LICENSE and src/public/thesaurus/LICENSE equal the upstream LICENSE.
+  NOTICE and README say Princeton terms, not MIT or Apache.
+- PROVENANCE.json hashes to e5a799a6, matching BUILD.json and PROVENANCE_SHA256
+  (build-thesaurus.mjs:9).
+- All 7 shipped outputs match BUILD.json bytes and sha256. Both converter hashes match
+  BUILD.json inputs.
+- The 147,306 vs 147,278 headword gap is recorded in BUILD.json, the manifest and the README
+  as sage-3972. I take that ruling as relayed.
+
+## Refusals
+
+The converter checks the provenance pin, the inventory, per-file hashes, byte totals,
+LICENSE equality and the NOTICE text (:21-40). Then it parses the record grammar with ASCII
+checks (:41-110), checks pointer targets (:129), checks the counts against EXPECTED (:132)
+and checks that the index matches the data (:133). It writes nothing before all of these
+pass (:155). The importer checks the archive hash and the provenance pin and validates every
+entry before writing. It refuses an existing destination and cleans up on a failed write.
+No runtime, route, CSP or worker change is in the diff.
+
+## Findings
+
+B1. The data test fails on any Node other than v26.8.1. build-thesaurus.mjs:147 writes
+`tools.node: process.version` into the build info. thesaurus-data.test.mjs:33 deepEquals the
+fresh info against the committed editor-data/BUILD.json, which says "v26.8.1". Any host Node
+upgrade, or a gate on a host with another Node, turns the webui suite red with no data change.
+The README (:30) states a >=24 baseline. Fix: keep the run's Node out of the compared
+object (record it separately, or drop it from the test's comparison) and leave the byte and
+hash comparisons as they are. Today's host runs v26.8.1, so Glomgold's 7/7 and 425/425 don't
+show this.
+
+C1 (gate). The queue gate says "named mutants killed". BUILD.md names no mutants, and G1
+doesn't name any. Reading the test statically, these guards have no test that reaches them,
+so a mutant removing any of them survives:
+- the NOTICE attribution check (:38);
+- the reverse index check (:100);
+- the pointer-target check (:129);
+- the counts-vs-EXPECTED refusal (:132). Test 2 asserts the literal counts, which catches a
+  converter change, but not removal of the refusal;
+- the normalized/index match (:133).
+Sage should name the mutants for the gate, or rule that none are required for a data row.
+At minimum I'd want the pointer-target and reverse-index checks covered, since "malformed
+record refuses the build" is the brief's acceptance.
+
+C2 (gate, not mine to decide). BUILD.md cites landing hold mos-1573: nothing lands until
+Jason decides on about 58 MB of permanent history (Vikunja 618). I found no ruling on refactor.
+Sage should confirm that hold is cleared before the landing gate.
+
+## Notes
+
+N1. Offsets of 10000000 and above are integer-like keys, so JSON.stringify emits them before
+zero-padded keys. The output is deterministic and the index carries the order, but a consumer
+must not rely on synsets key order.
+N2. THESAURUS_ARCHIVE gates the round-trip test (skipped otherwise). Glomgold's final logs
+show it ran with 0 skips. Sage's gate should set it too, or the run silently drops the one
+check that ties the vendored text to the archive.