docs: a tested dependency census for the backlog, and #1250's measured numbers (BACKLOG #1250)

[wshallwshall]

What this is

A committed, tested dependency census for docs/BACKLOG.md, plus the measured numbers written into BACKLOG #1250.

THE BACKLOG WAS NOT MOVED, AND THE MOVE IS STILL BLOCKED. #1250's precondition is vault-side (atomic allocation, a high-water ratchet, installed hooks -- the Q5 ruling of 2026-08-13) and is unbuilt. The destination is still the weaker repository. Nothing here changes that, and the new tool moves nothing.

Why

#1250 has carried a hand count three times and it drifted between each: 66 tracked referencing files at filing, 78 at the 2026-08-20 re-score. A cross-repo migration steered by a number nobody can re-derive is how allocation collisions come back silently, which is the defect the ledger gate exists to prevent.

What was built

scripts/docs/backlog_dependency_census.py walks the tracked corpus (git ls-files, or git ls-tree at --ref) and classifies every dependent two ways.

By role, so the four pieces of machinery the item names are findable by name: pre-commit gate, pre-commit config, status checker, citation checker, CI workflow, allocator, test, CI config, tooling, doc, engine, root.

By mechanism, because a path literal and a parser import fail differently under a move:

mechanism	count at 975c13c1	how it fails
path-literal	64	the open fails, or silently reads nothing
path-pattern	3	a regex or glob spelling; invisible to a literal search
parse-items	31	survives a file move; breaks on a DEFAULT_SOURCES change
bare-filename	95	prose or a relative link
markdown-link	30	a dead link a link checker reports
archive-sibling	26	also names the archive half of the namespace

A file carries a set and the sets overlap, so those sum above the 100 total.

Two findings a better grep would not have produced

1. .pre-commit-config.yaml depends on the ledger through a REGEX. Its backlog-parses hook is scoped files: ^docs/(BACKLOG\.md|archive/backlog/.*\.md)$. A grep -F docs/BACKLOG.md over that file returns exactly one line -- a prose comment -- and misses the functional filter. backlog-hygiene.yml carries the same shape in its own diff check. This is the shape of false zero the item exists to prevent, and it is pinned by a test asserting BOTH halves: the literal search misses it, the census does not.
2. Five dependents name no path at all. scripts/coord/landed_citation_screen.py, scripts/coord/throughput.py, scripts/quality/expiry_audit.py, tests/test_dangling_citation_check.py, tests/test_gate_ci_mirror_parity.py read the ledger through parse_items. Re-pointing them as path literals would fix nothing.

Also recorded in the item: the filed blast-radius block's own eight directory buckets sum to 57, not 66. Nothing was wrong at the time and nothing reported the gap, because a hand count carries no control.

One false positive, found and recorded

The path-pattern detector originally accepted a bare BACKLOG*. Listing the hits rather than reading the count found docs/benchmarks/HANDOFF-enginebox-step2-step3.md -- prose reading "pages scale with BACKLOG", where a closing bold marker is a star. One wrong hit in a population of four is a quarter of the one class this tool exists to surface, and it would have gone into the item as a migration target.

Fixed so every arm has to reach the md, with both directions pinned: three prose shapes that must NOT fire, three real spellings that must. The finding is written into #1250 rather than fixed out of sight, because "found by listing the hits, not by reading the count" is the transferable part.

Measured numbers, now in the #1250 body

what	filed	re-scored	measured at 975c13c1
tracked files depending on the ledger	66	78	100
of those, test files	13	15	29
open items with "ASVS" in the heading	103	106	107

Positive controls

Seven must-fire controls name a file and the mechanism it must be found under. Three must-not-fire controls name ordinary tracked files that must carry none. A control that does not fire exits 2 and prints the counts as untrustworthy rather than as fact.

The historic arm is the strongest evidence available. Run at c2241cfe -- the ref #1250 was filed on -- the census returns 103 ASVS headings, the number a human wrote into the banner that night, from an independently written instrument. That run also legitimately misses one control: at c2241cfe the backlog-parses hook (BACKLOG #1259) had not landed, so .pre-commit-config.yaml carried no ledger dependency yet. The script says so explicitly rather than letting the exit code read as a broken detector, and a test pins that exact behaviour.

tests/test_backlog_dependency_census.py reds when a detector goes dead and when one goes generic. Either arm alone passes against a broken instrument, which is why both are there.

What the census cannot see, stated in the tool and in the item

• The vault -- the migration's destination, and the whole of the unbuilt precondition.
• Anything outside this repository: other clones, the published history, sdists on PyPI, branch protection.
• Untracked and ignored files; a path built at run time; a dependency that never names the file.
• The parked deficits, and I am not guessing at a number. The 2026-08-14 ruling parks a coordination-tooling or seat-topology item outside the public ledger, and with no private ledger those land in <git-common-dir>/mefor-coord/. That directory is untracked, machine-local, and carries no machine-readable marker separating a parked deficit from any other handoff note, so no count off it is reproducible on another clone. The population is known to be at least one -- the ruling's own instance -- and is not otherwise quantified.

Assumptions taken, since a Builder gets one turn

• The score line and the banner's own historical counts were left exactly as written. The brief says do not re-score, and the item itself says the priority line predates the 2026-08-14 ruling and is left unchanged deliberately. Editing the counts inside a dated re-score rationale would rewrite a record rather than supersede it, so the new measurement is a dated body paragraph that names which three figures it replaces. If the owner wants the banner and the priority-table row rewritten in place, that is a one-line follow-up.
• archive-sibling alone does not count as a dependency on docs/BACKLOG.md. One tracked file (docs/adr/0076-...md) names only the archive half; it is reported on its own line rather than in the headline, because counting it would answer a different question. The two ledger files themselves are reported as SUBJECT, not as dependents -- including them inflated every total by two.
• The census is not wired as a gate. Only its test runs in CI. Making the census itself a required check is a separate decision.

Checks

Interpreter verification, per the brief -- it imports from this worktree:

$ /c/Users/Scott/Code/MessageFoundry/.venv/Scripts/python.exe -c "import messagefoundry; print(messagefoundry.__file__)"
C:\Users\Scott\Code\MessageFoundry\.claude\worktrees\agent-a3d383e125d7303dc\messagefoundry\__init__.py

Run and green:

• ruff format --check + ruff check on both new files

• mypy --strict on scripts/docs/backlog_dependency_census.py (CI's mypy scope is messagefoundry + messagefoundry_webconsole, so this is beyond the gate)

• scripts/hooks/ledger_check.py -- exit 0

• scripts/docs/backlog_status_check.py -- OK, 664 items, each declaring exactly one status

• scripts/docs/backlog_citation_check.py --base origin/main --head HEAD -- OK, 0 citations in scope

• pytest tests/test_backlog_dependency_census.py -- 23 passed

• pytest tests/test_tooling_partition.py tests/test_cp1252_console_safety.py tests/test_script_root_anchoring.py -- 47 passed

• every pre-commit hook, on both commits (never --no-verify)

• pytest tests/test_backlog_status_check.py tests/test_backlog_citation_check.py tests/test_dangling_citation_check.py tests/test_ledger_check.py tests/test_link_resolution.py tests/test_tooling_partition.py (with the census test) -- 182 passed

Not run: the full suite (BACKLOG #1014, port collision).

Legs that must be read after this process is gone: the tooling job on both OS legs (the new test is registered in tests/tooling_manifest.txt), backlog-hygiene, and the CI gate. gh pr checks renders CANCELLED as a fail -- tally the check-runs API instead.

Commits

Four, one layer each, and the last touches docs/BACKLOG.md alone. Commits 3 and 4 are the false-positive fix and the figures it corrects; they are separate commits rather than an amend because the branch was already pushed and a force-push is not recoverable.

🤖 Generated with Claude Code