docs(backlog): close six rows whose every limb is verified landed

[wshallwshall]

What this closes

Six rows whose every limb is verified landed: #1272, #1291, #1345, #1366, #1367, #1410. Open rows go 279 to 273.

Why six and not 164

A complete run of scripts/coord/landed_citation_screen.py over all 278 open rows flagged 164 — 145 CITED, 19 MENTIONED — with its built-in controls passing.

That is a read-before-you-dispatch signal, not a completion claim, and the tool refuses that reading in its own words: landed code can cite a row while filing it, while testing around it, or while recording why it was not done.

Narrowing to rows that are CITED and carry a merged work PR and have no open PR and score value 1–2 gives 19 candidates.

Each of the 19 was proved individually, defaulting to refuse

The asymmetry drove the default: refusing a correct flip costs one re-read later, while accepting a wrong one removes the row from view, so the remaining work is never done and nothing reports a problem.

Six passed. Thirteen did not, each naming a specific unlanded limb:

Row	What is still missing
#1020	three of its own "NOT BUILT" limbs unlanded; ADR 0167 still Proposed
#1061	the row's own banner says it must not close, and two merged PRs agree
#1282	the doc correction the row calls "what is left" has not landed
#1332	no limb landed at all — it is a tombstone for #1086
#1357	the decision record still states the retired value
#1361	the 25 budgeted message sites are not rewritten
#1368	branch-leak-scan.yml pins the same floor with no freshness check
#1372	the pinning test the row names as its only remainder never landed
#1383	45 assignment-shaped citations of refused keys remain
#1404	the enforce_admins claim is still wrong in four places
#1405	the second half of the two-step never landed
#1409	names two gaps that have no other number, so closing loses them
#1406	the injection never reaches the spawn path the Console actually uses

#1409 and #1368 are the instructive refusals. Both had every stated limb landed, and both were refused anyway because the verifier found something the row still uniquely records. That is the check working.

#1406's re-run has since completed, and it is a refusal too — 1 of 3 limbs. The headroom hook is wired on matcher ^(Task|Agent|Workflow)$|spawn_task (.claude/settings.json:60), but CLAUDE.md:290-292 says a Console spawns a Builder through a Bash(claude:*) grant, which no tool-name matcher can select. The shipped contract test does not merely fail to cover that path — tests/test_claude_settings_contract.py:151 puts Bash in _NEAR_MISS_TOOLS and asserts no matcher wiring this hook may select it. So the largest spend in the system, launching a whole session, still gets no reading, and nothing records the exclusion.

The verifier drove it rather than reading it: a Bash/claude --bg payload piped into the hook produced no output and exit 0, while a Task payload produced a full reading.

So the final count is six flipped, thirteen refused.

Shape of each edit

The check-mark banner leads. The original open banner is demoted, not deleted — its glyph is stripped and its filing prose stays beneath, which is the shape #1439 and #1432 already carry on main. The score line is untouched.

Exactly one status banner per row, asserted from the written file rather than from the plan.

Checks

• backlog_status_check.py green: 670 items, each declaring exactly one status.
• All six parse as closed, with one closed banner and zero open banners.
• 113 tests pass across the status, citation, ledger and link-resolution legs.
• Additive shape: 18 insertions, 6 deletions, docs/BACKLOG.md only.

Not enqueued, auto-merge not armed.

🤖 Generated with Claude Code