stack/agents/researcher/work/queue-80/r2/BUILD.md

Row 80 (#1574) round 2: mechanical rebase onto row 93

Owner: Researcher. Reviewer: Filbert (approved round 1, comment 28188, on manifest base 57a2ddc5). Row 93 landed as 4da68152 and changed packages/bus/src/broker.mjs and packages/bus/README.md, two of the three candidate paths, so the round 1 manifest hashes no longer match. Decision 127: a new round. This round changes no code, test, README sentence or mutant.

Base and method

• Base da6f1166d63f8047f8b59c9742815c63982f8ae9 (base.txt), a fresh detached worktree (/mnt/storage/scratch/researcher/q80-r2).
• git apply --check --index r1/build.patch then git apply --index r1/build.patch: clean, no fuzz, no hand merge.
• build.patch is git diff --cached da6f1166 of the same three paths.
• candidate-manifest.sha256: the same three paths, repo-relative, verifies with sha256sum -c from the worktree root.

path	round 1 (base 57a2ddc5)	round 2 (base da6f1166)
packages/bus/README.md	02fae293...	675dbeca...
packages/bus/src/broker.mjs	f685af1c...	3a9de6f0...
packages/bus/tests/agents.test.mjs	076137bc...	076137bc... (same)

r1-to-r2 diff of the patch

The + and - lines of the two patches (file headers excluded) are 117 each and identical (diff of the two filtered line lists is empty). Every context line is identical too. The two patch files differ only in the index lines and in the hunk start lines:

file	hunk	r1 start	r2 start	shift
README.md	1	108	129	+21
README.md	2	127	150	+23
README.md	3	149	172	+23
README.md	4	159	182	+23
broker.mjs	1 (the STRANDED constant)	43	43	0
broker.mjs	2 (stranded query, .map)	1039	1213	+174
agents.test.mjs	1	55	55	0
agents.test.mjs	2	157	157	0

agents.test.mjs is byte-identical to round 1. Row 93 added about 174 lines of its own above the stranded query in broker.mjs and 21 to 23 lines above my README hunks; none of it overlaps mine.

Mutants

mutants.py is byte-identical to round 1 (cmp). All 15 anchors match exactly once in the r2 broker.mjs. The line numbers in the mutant table's labels are round 1's (the runner matches text, not lines); on this base the stranded query is at 1222 instead of 1048, M12 at 1227, M13 at 1106 and M15 at 1226. The "expected killer" test lines are unchanged because the test file is.

Runs

Run on 2026-10-13 (15:54Z onward), in the slot Sage gave for this round, after a cwd-scoped busy check that read 0. The manifest was verified before the runs and after them (all three paths OK). Logs are under out/.

• Bus suite, node --test 'tests/*.test.mjs' in packages/bus, run in its own session with the group killed afterwards, DOCKER_HOST=unix:///nonexistent.sock (out/bus.log): 127 tests, 127 pass, 0 fail, 0 cancelled, 0 skipped, no ✖. Round 1 had 120; the extra seven are row 93's tests on this base, not this candidate.
• mutants.py <r2 worktree> r2 (out/mutants-run.log, out/summary.txt, one log per mutant): baseline green, 15 of 15 killed, exit 0. Manifest clean after the last restore. No process had a cwd in the worktree afterwards.
• The test lines that fail for each mutant are the same as round 1's (out/M*.log against r1/out/M*.log, the agents.test.mjs:NNN sets compared per mutant, equal for all of M01 to M11 and M13 to M15). M12 is flagged "error marker in log" by the runner exactly as in round 1: it removes stranded from the claim row. The key-list assertion at agents.test.mjs:58 fails, and the stranded tests throw TypeError: Cannot read properties of undefined (reading length and map), which is the marker. That is a real kill, not a crash of the runner.
• Not run here: the other bus-adjacent suites and the scripts/test-*.sh. Only packages/bus paths changed against the approved round 1 content, and none of the diff beyond hunk offsets is new. Sage's landing gate reruns them in any case.
• The worktree is kept for the review (/mnt/storage/scratch/researcher/q80-r2, three paths staged, no node_modules link needed by the bus suite).