Appearance
Later — deferred, each with a moment that reopens it
This project's own deferred list, kept by the rule it gives everyone else: a revisit trigger that is a moment, not a date. A deferral with no trigger is an abandonment wearing better clothes, and this file exists so the ones here can be told apart from the ones nobody wrote down.
What belongs here: something real, understood, and consciously not done now. Not a wish, not an idea — those go to the backlog or nowhere. What does not belong here: anything that stops work now, which is a task with a blocker instead.
./diagrams is at its budget
./diagrams is at its budgetMeasured 2026-08-15: 499 of a 500-line chapter budget, while self-maintenance §What-a-capability-owes requires a diagram per new mechanic. The question was never where a diagram goes but whether diagrams is a chapter or an index — answered 2026-09-06 by use: two releases running put their diagrams beside the rules they draw, and twice in one session a repair owed none and said so. The gallery reading was the collision course, not the budget: a file holding every diagram meets 500 at each new mechanic; one holding only what crosses several rules does not.
Measure what the day-one cut actually bought — run 2026-08-01, partly falsified
Closed by measurement, and the verdict is split. The ordering claim confirmed: the first task is now written second, before every document, where it used to arrive after the scaffolding. The volume and time claims falsified against criteria written before the run and kept in runs: still ten to thirteen files, and turn two still at four hundred seconds against a stated threshold of three hundred. runs carries the table.
What it taught, and the next thing to try: the gate could see order and enforced it; it cannot see emptiness, so once the task exists every document passes. The corpus says day one is four things and the hook enforces something narrower — the gap between them is the thirteen files. The predicate that would close it is refuse a document whose body is a heading and a template's braces: checkable, a commission, and the shape that has worked twice. Not attempted, because the criterion comes first — and that half is open work under a struck heading, which the 2026-09-06 audit read past. Its form question is the Nothing here measures whether a DOCUMENT works entry, which asks the same thing.
The criteria that made the verdict possible — written before the run, because criteria written after a result are not criteria — are in runs under the day-one cut.
A generated Opportunity-Solution-Tree view over the specs
Deferred because the data half landed on 2026-08-06 — the spec's Opportunity field names the need and cites its insight, so every new spec feeds the tree — and only the generator remains: mermaid between markers (patterns §6, same shape as touched by:), a day of work with no requester yet. A view nobody reads is furniture; the field is not, because a solution with no opportunity above it now fails a read even without the picture.
Revisit when the first project running product discovery asks to see the tree, or a review finds a solution shipped with no opportunity above it.
A terminal UI over _ops/ — deferred, with the boundary that would shape it
Named 2026-09-07, when the owner asked how hard a herdr-like terminal UI would be. The deferred thing is the build; what follows is why not, and what is decided so it is not re-argued.
Native-first, dated. herdr (read 2026-09-07) supervises CLI agents rather than being one, so this skill already runs inside it and there is nothing to port; its plugin surface carries executables, not context.
The boundary, stated without naming anyone. A pane holding a live process belongs to whatever supervises the session. This system refuses that half on purpose — nothing load-bearing lives in a session — and owns the durable one: who exists, what they may do without asking, what is assigned, what it costs, what waits on the owner. Build the session half here and the result is a worse supervisor bought with the founding premise. The move the other way is cheap and holds for any supervisor: it sees that an agent stopped and not why, while _ops/ knows which task, which gate and what the wait costs — the reason can go on someone else's status row without this system growing a session model.
Settled, each an existing rule applied: no model in memory, every repaint re-reads · no silent subset, a count line says what it could not read · one writer — it reads, the agent writes, the owner edits · no chat outside a thread · the first view is the bill for the owner's attention, which is /status repainting · not columns, because a wave is not a column.
Why not now. Nobody asked for the build, and the one screen with no substitute — what changed in _ops/ while I was away — has never been wanted aloud.
The candidate form: the parser refuses a malformed entity and the count line reports it. The mutant is a tree where one task's status line is bolded differently and the display shows the rest without a word; the twin is that tree read whole. scripts/transition.py's field() already reads the prose form tolerantly.
Revisit when an owner asks what changed while I was away and answering means reading git log -- _ops/ by hand.
starts: webhook — an external trigger the automation can declare
Reopened as a live candidate 2026-08-06 — the listener now exists somewhere real. OpenClaw was measured on this machine: a resident WebSocket gateway, cron, agent hooks — the first runtime that could actually hold a declared trigger, where every earlier survey found only serverless sessions. The shape stays settled: webhook:<name> beside schedule:<cron>, executed by the host that has a listener, under the schedule's own law — one that silently did not run is worse than one that says it was late.
Revisit when the first automation actually needs an external event — and the first step then is verifying OpenClaw's trigger surface (its cron/gateway API), not building one.
A tripwire for work bypassing the machinery wholesale — still open; a neighbour of it shipped
The evidence: a live project split into two eras at one commit — before it, six honest run records and twelve tasks with done-definitions; after it, 0 of 33 commits touched _ops while the site shipped, decisions changed unrecorded (domain, language count, act format), and the board froze at "created". The guard fired on every commit and had nothing to say, because no check reads work advancing while the record stands still. The shape of the form: the project preflight warns — never fails — when the last N commits carry zero _ops/ paths in a repo whose guide declares an operator. Warn, not fail, and once: site work is legitimate; the record going dark is the smell. Revisit when: the next takeover audit or field report shows a second project with a dark era — twice is the threshold everywhere here.
Read again 2026-09-16 against a second live project: the trigger did NOT fire — 12 of its 15 commits touch _ops/. What that audit found was a neighbouring shape with its own measurement, 4 of 15 commits moving files outside _ops/ with nothing under _ops/ in them, and the guard's §24 holds that one. The two are not the same rule — §24 reads one commit and its parent, this entry reads a run of N — and a form built for a trigger that has not fired is a rule built on one project's habits.
A channel that reaches the owner who is away
Requests age in files, and nothing tells an owner who is not at the terminal. Evaluated 2026-08-14: Agent-Reach (71k★, MIT) — declined; despite the name it is inbound only (agents reading platforms via cookie scrapers, ToS-gray, churn-prone), it contains no notification surface at all. The honest candidates remain the harness's own push notifications and a Telegram-bot shelf row (one token, owner-controlled, off-terminal). The row exists since 2026-09-25: Apprise (BSD-2-Clause, one CLI for ~100 destinations, Telegram among them) — catalogue → A channel to the owner who is away. What stays deferred is wiring it in, and the trigger for that has not fired: a shelf row answers which tool, not whether the need is real. Revisit when: a request ages past its threshold unseen in a real project — that is the moment the need names the wiring.
A market-recon distillate template for research/
The folder exists with no named shape (project-layout — "studies and their distillates"), and a live project's recon produced clean per-study documents with nowhere agreed to converge. The substance worth keeping from the 2026-08-14 gap analysis: analogs land in the competitor register (now with stage/outcome) · what-to-borrow and risks-and-failures as distillate sections citing register rows · the delta is the positioning brief, homed in BRAND-template.md §Positioning. Revisit when: the next real market-recon study is commissioned — a template written before its second use is a guess.
The advisor-side half of the doors: nothing forces wiring at stand-up — built 2026-09-17
Measured twice, which is the threshold. N89 at 0 of 3 (2026-08-14): every graded day-one run stood the project up ad hoc, so the guard's doors check never got its chance. The second measurement came from the field (2026-09-16): a live project, stood up and operated by this skill, held the guard in _ops/scripts/ one version behind with .git/hooks/ holding only git's samples — every gate a description, and underneath it nine tasks in two shapes, two accepted without review, an assignee in no file.
Built: the SessionStart hook runs the same scripts/guard-status.py a person runs and states what it found, once, as a fact and never as an order. Two silences are load-bearing and tested — no guard at all is the arriving flow's debt-list item and is visible without a hook; a project already wired owes no line, and that is the assertion its mutant fails. test-audit-gate.sh 128/128.
The council has a field but no gate, and a document has no seam to hold one
This is a debt against the capability bar, not work consciously deferred — the distinction this file's own contract draws. It stays because the missing half IS deferrable and its trigger is real; what belongs beside the bar, and not in a list of things not yet built, is the admission that the bar was not met.
Named 2026-08-14, at the release that shipped it. The capability bar has four clauses and says no exceptions for small ones; the council shipped with the showcase trio and a declared field (angles · voices · provider), and without the other two — no form where it can fail, no mutation test, and therefore no dated measurement of the field ever appearing. This is written here rather than implied met.
Why no form today. A council is a consultation: zero standing footprint, so its synthesis is an answer in a conversation, and there is no artifact for a gate to read. Every enforceable rule in this corpus keys on a file — a staged commit, a row, a request. The declaration is real guidance and unenforced guidance, which is the class this corpus keeps measuring at ~0/5.
Revisit when either becomes true: a council's synthesis is written to a file (the owner asks to keep it, or a research flow persists one) — that file is the seam, and the gate is a missing-declaration refusal with its mutant · or the next full round measures whether a council fires with its price and its declaration at all, which would give the claim its date. If a round shows the field is simply not written, the honest repair is to delete the sentence rather than repeat it louder.
The repair exemption is judgement, and the test that would replace it has not paid its price
Named 2026-09-07, the day it was written and taken back out. self-maintenance §4 says a repair owes only the fact, not the diagram and not the situation. Nothing decides what a repair is — clauses 1–3 of the bar name artifacts a script can find; this one is answered by a sentence in the changelog, written by the person who wants the answer, and it was claimed twice in one session before anything checked it.
The test that would close it: a change shipping a new form — a gate, a field, a refusal — is a new mechanic and owes the trio; a change that only alters what an existing form reads is a repair. Read off the diff, not off the author's account.
Why it did not ship the day it was written. self-maintenance's promotion ladder puts a rule at rung 3 and charges a week between the first dated line and the promotion, and it still reproduces; this was conceived and written into a standing chapter in one sitting, which is that section's own dotted branch — promoted the same day → a rule that was true once, on one machine, about one version. Its exception is a defect with a live blast radius, and a test telling authors how to classify their own changes is a lesson.
What a form would read, when it has earned the rung: preflight.sh §1b-bis already resolves the last tag and names ${last_tag}..HEAD as the lens range, and §11 already reads the entry being shipped. A check greps that range's diff for a new section header in templates/company-preflight.sh — the shape that file writes is # N · , never §, whose occurrences there are all prose cross-references — or a new required field in a template, which is what +| **Verdict** | was.
The pair only works if the check refuses a form signal answered by an unbacked claim — that is the condition, and a trim once replaced it with the weaker "the twin must pass", which states what lets the honest case through and not what stops the dishonest one. The mutant is a release whose diff adds a section header and whose entry says repair with no trio; the twin is one whose diff adds no form and says the same (a first draft accepted the word repair as an answer, so the mutant passed) — the twin must pass, or the check refuses everything.
Revisit when the exemption is claimed in a release at least a week after 2026-09-07 and still reads right. Do not count the citations inside the release that wrote it: 0.2.16 applied the idea three times in one commit, which is one occasion.
The two capability bars have diverged, and neither repository says so
Named 2026-09-07 by a lens asked to compare them. self-maintenance's bar and the sibling's AGENTS.md bar agree on clauses 1 and 2 — a form where the rule can fail, a mutation test — and are disjoint on 3 and 4: this repository asks for the claim dated with its measurement and the showcase trio, the sibling for a scenario and the fixture it needs and a door. Each file says four things as though its own partition were the finding.
It is mostly one set cut two ways, and that is the part worth fixing. The sibling's door is this repository's rule too — AGENTS.md carries every capability has a door outside the bar. The sibling's scenario and fixture is enforced here in code and omitted from the stated bar: preflight.sh warns when a version bumps and evals/runsheet.tsv is unchanged. The trio lives in the sibling as a session-loop step with a narrower fact rule; the dated claim is folded into its clause 1.
Why not now. Reconciling two bars is a change to both repositories' always-loaded contracts, and the honest first step was the smaller one already taken: §4 no longer claims both siblings' ledgers carry the receipts for clauses the sibling does not hold. Merging the partitions is a bigger move than one release, and doing it while the divergence is one day old repeats the mistake the entry above records.
Revisit when a third repository takes this bar, or when a capability ships in one repository and is ported to the other — the first time the two lists have to agree on one object, which is when a difference stops being a wording question.
The promotion ladder's week has no form, only a citation a form could read
Named 2026-08-21, at the release that shipped it. self-maintenance now says a lesson waits a week between its dated field-note line and its promotion to a rule — a filter against a corpus that fills with laws written the afternoon something broke. Two of its three parts are already machine-readable and one is not: field notes carry a Date column, and a promotion is told to cite the line it came from, but nothing reads the citation against the note's own date, and nothing notices a rule that cites nothing at all.
Why no form today. The seam exists — a promotion touches a rule file and a field-notes line in the same commit — but the gate would have to tell a promotion from any other edit to a rule file, and there is no field declaring one. Inventing a field for the gate's convenience is the shape this corpus refuses: the form should read something the work already writes.
The precondition landed 2026-08-27, and the gate deliberately did not. Read the trigger below before concluding otherwise: the column existing is not the trigger, a closure written by real work is. The closure this entry waits on (
checking) was prose with no cell to hold it; it is a column now. Building the gate in the same hour would be the law written the afternoon something broke, which is the exact thing this ladder exists to filter.
Revisit when a Closed cell is filled by a sweep closing real work in a project running this skill, and that closure reaches here as a field report — not by the person who added the column, and not by a note written to demonstrate it.
The template can hold the closure now; what the gate waits on is one written for its own reasons, because a gate built on a shape nobody has used yet is a gate built on a guess. The gate then reads: a rule added to an always-loaded file in the same commit as a closure whose dated origin is under seven days old is refused, naming both dates. The mutant is a same-day promotion; the twin is one dated eight days back. Until then the week is judgement, and it is written here rather than believed.
Depth was demoted before its tag, and owes the measurement it never had
Named 2026-09-10, by the adversarial lens on the release that shipped it. A three-rung depth ladder for research — orienting · deciding · standing — went in as a required field with a guard, and no measured defect stood behind it. The bar asks for the claim dated with its measurement; the date belonged to a runtime check and nobody had been observed getting it wrong. A rule that cannot point back at its dated origin has not climbed — it was declared.
Demoted the same day to guidance that admits it. Two defects went with the gate: its standing cross-check guarded on sources, a path no operated project has, so the clause the entry led with was dead everywhere it shipped; and its enum accepted the empty string, waving a blank Depth: through while refusing an honest Depth: TBD.
Revisit when a research request is measured against a stated rung, or an owner objects to what a search cost after it ran.
The lens round has no form, and its range cannot see the repairs it prompts
Named 2026-09-07, by the adversarial lens reading the release that added the repair pass. Nothing counts rounds, so a skipped lens and an empty one are the same artifact — preflight.sh says so at every release, which is honest and is not a check.
And the range has a hole the moment the tag is cut early. <last tag>..HEAD contains a repair until that repair's own tag exists, after which it is an ancestor and every later range excludes it. Measured: 1727542, the repair 0.2.16's re-run produced, is an ancestor of v0.2.16 and absent from git log v0.2.16..5f76d30 — the commit that motivated the rule is the one a deferral to "the next round" would have missed. lenses closes it by ordering, which is a rule and not a form: nothing refuses a tag cut between rounds.
The candidate form: record the reviewed range in the entry, and have preflight refuse a tag when that range does not reach HEAD. The mutant is a release whose recorded range stops short of its own repair commits; the twin is one whose range reaches them.
The sibling already ships two thirds of this. Its scripts/lens.sh <lens> <base-ref> runs one lens as an isolated headless session — own config home, no connectors, Write denied — beside a run record that lists not run rather than omitting it. And it has the state this project lacked: a lens that exhausts its budget is recorded as not completed. That half was ported 2026-09-10 after nine agents died on usage limits across three rounds with nowhere to record it but prose.
Why no form today. A runner is a capability and owes the bar. And the cost is undeclared — one round per release, with no number on a lens pass anywhere, which is the adversarial lens's own fourth question unanswered by a rule it reviewed.
Revisit when a release ships a defect the previous release's repair introduced, or when any lens round is measured for what it costs.
Convergence between independent reviewers is not read as a signal
Named 2026-08-28, from qa-swarm — surveyed in the 0.2.14 entry, whose catalogue row 0.2.16 removed. The four lenses already run independently and each reports even when empty, which is half of what that harness does. The half we do not do is the cheap one: when two of them land on the same file and the same line, that agreement is evidence, and this project currently treats it as a duplicate to be merged away. A finding two blind readers reached separately is not the same object as a finding one reader reached twice.
Why no form today. Lens output is prose in a task notification, not a record with a file and a line — so there is nothing to compare. The same objection LATER.md raised about the contradiction stop applies, and the same answer resolves it: wait for a flow that wants the finding written structurally for its own reasons rather than adding the shape for a counter's sake. The review request is the candidate — a review that lands as findings with locations is useful to the requester whether or not anything counts convergence.
Revisit when a review answer carries findings as rows. The gate then reads: a finding two independent reviewers reached is promoted, and the promotion is recorded, so a later reader can tell an agreed finding from a repeated one. The mutant is two findings from ONE reviewer merged and promoted; the twin is the same pair from two.
An install copy carries what the source gitignores, and the inventory reads it as ok
Measured 2026-09-07, during 0.2.16's re-sync. Both copy-route installs on this machine hold 13 MB against a tracked source of 3.1 — ledger/, brandkit/ and .claude/worktrees/, all gitignored here and shipped by nothing (git ls-files over them returns zero). An earlier rsync ran without excludes and every re-sync since preserved them.
Why it is more than megabytes. An install directory holding a lens worktree holds a second checkout of this repository, at some other commit, inside a tree a runtime loads. Nothing reads it; nothing says it is there either. find-installs.sh reads both copies as ok because it checks the version stamp and not what else came along — the same shape as the mount-config entry, where the inventory is silent about a state it never looks at.
Half of it is closed: the copies moved this time by git archive <tag> | tar -x, which emits tracked files only and cannot carry an ignored path — the drift stops growing, and does not shrink.
And one act is owed rather than deferred: the 13 MB stayed because the comparison CLAUDE.md requires — untracked artifacts checked against origin/main before removal — was not run. That waits on somebody, not on a moment. Do it the next time a copy install is touched for any reason.
Why no form today. The obvious check — git check-ignore the copy's paths against the source — needs the source present, which is the one thing a copy install does not guarantee.
Revisit when a copy install is moved by anything other than git archive, or a second machine shows the same spread.
The inventory is silent when a mount's CONFIG goes and its directory stays
Named 2026-09-05, doing the ritual. find-installs.sh gained a hermes row because a mount is an install even though no file lands anywhere, and its comment covers one direction: "it outlives the directory it points at, which is how a cleanup elsewhere silently unplugs it." The other direction happened this release. ~/.hermes/ was removed between two runs; the row simply stopped appearing, and ~/.agents/skills/opsinist/skills still holds 21 files that hermes had been mounting. The inventory's output is the canon by this project's own rule, so a canon that shrinks without saying why is a canon that has to be re-derived by hand — which is what it exists to stop.
Why no form today. The honest check is not "did a row vanish" — rows vanish legitimately all the time, and a diff against a remembered count is the shape this corpus refuses. What it wants is a directory that looks mounted with no config naming it, and "looks mounted" has no test yet beyond "a runtime we happen to know about used to point here". Guessing at that would be a check that cries wolf, which is worse than this silence.
Revisit when a runtime's mount is recorded somewhere the inventory can read back — a stamp the install itself carries, rather than a config only the runtime owns. The gate then reads: a directory carrying our manifest, inside a runtime root, that no config mounts is reported as orphaned. The mutant is a live mount reported as orphaned; the twin is this exact case — config gone, directory full — reported once and by name.
A "should this exist at all" ladder, and whether it changes anything
Named 2026-08-22, from a third-party plugin the owner asked about (DietrichGebert/ponytail, MIT, read 2026-08-22). Its content is one ladder an agent walks before writing code: does this need to exist · is it already in the codebase · standard library · a native platform feature · an installed dependency · can it be one line · only then a minimum implementation.
Why it is a real gap. choosing-tools holds a ladder for which vendor. Nothing shipped asks should this code exist at all; the nearest thing, native-first, lives in AGENTS.md, which governs this repository and never reaches a project.
Why it is not taken. Its headline rates carry no denominator, no corpus and no date, and 100% safe is not falsifiable — the ladder can be taken, those numbers cannot be quoted. And a finding does not become a rule the day it is found: two capabilities shipped in 0.2.9 whose prose measures 1 in 10, so a third before the first two are measured is what self-maintenance refuses.
What the round bought, and it narrows rather than answers. The prose arm measured 1 of 10. N102 is N97 with "then commit it" in the turn, the only difference, and it converted 3 of 3 — the two rates do not share a denominator: 1-of-10 is over every dispatch, 3-of-3 over the three of five that reached the gate. So a ladder shipped as prose lands in the band this corpus measures near zero; whether one shipped as a form changes anything is untouched. If it arrives, it arrives as a form.
Revisit when N97/N98 are re-run against a decision this round left open — whether a scenario meaning to reach a gate must say commit it in the turn, or whether not-committing is itself the behaviour under test. The old trigger said "when N97/N98 have measured the gates in the wild", and they ran without measuring them: a validator reaches only a worker that commits, ten runs committed nothing, and the round closes with "0.2.9 has no gate measurement".
Do not install that plugin alongside this skill — two plugins instructing one agent about how to write code is the shadowing trap runs measured.
Closed 2026-09-06 and reopened the same day by three lenses. The closure called the round's prose arm "a declared gate, in the field" and cited a page whose closing line says the gates were never reached. An audit that reads triggers against reality has to read the evidence file, not the trigger's summary of what would satisfy it.
Nothing here measures whether a DOCUMENT works, only whether a rule holds
What it is. Every scenario in evals/ puts a run in a situation and counts whether the rule survived. None of them asks the other question: given this document and a task, what comes back — how many tries to a clean result, how many violations of what it teaches, how big the output, how far each iteration moved, what the reading cost. A rule can hold inside a document nobody can use, and a document can produce good work while the rule inside it is never reached. The corpus has sixty-odd documents and no measurement of any of them as a document.
Why not now. It is a second harness, not a scenario: several runs against one task, an artefact compared across them, and counts that are not pass/fail. The existing rig measures a verdict; this measures a distribution. Building it while the scenario rig still owes RUNS — N97 and N98 have never measured the gates they were written for — would leave two half-harnesses instead of one whole one. (An earlier draft of this line said those scenarios owed fixtures; all four named are bound in evals/runsheet.tsv, and two of the ids were written without their N.)
What reopens it. A document is rewritten for readability and nobody can say whether it got better — the argument that has already happened twice about the always-loaded core and once about diagrams. That is the moment: the first time a rewrite is proposed and defended with taste alone.
Where the method came from. Sanity Labs measured their own design system this way and published the findings, the method and the runner (catalogue, Does the design system survive an agent?). Their counts are the borrowable half; their tool is theirs.
A lens's verdict never says how it looked — and three times it looked at a proxy
Three false verdicts, two shapes, one cause. Two absences on 2026-09-10: "the 500-line budget has no form" — the lens grepped 500, while preflight.sh reads chapter_budget — and "use-cases has no situation for the contradiction stop", where it grepped contradict in a file whose rows are the owner's words by construction. And one confirmation on 2026-09-11: three of four lenses agreed that check-structure fails a drifted count, each having read the script's fail() call, while preflight.sh §10 renders every FAIL: it prints as a warning, never fatal. No lens caught it.
Each read a proxy for the thing — a name for a mechanic, a producer for its consumer — and independent readers do not help when they share a method: three lenses, one path, one wrong answer, in a round that looked clean. The form this wants is the cut Reads against makes between none found and not checked: a verdict states what was checked and where, so "it calls fail()" reads as the half-check it is. Home: lenses, What the lenses do not cover.
Rung 2, not a rule — three occasions in two days, and rung 3 costs a week (self-maintenance, the ladder). Revisit when a fourth proxy verdict is caught, or the week has passed and these still reproduce. The cost is real: every finding would carry the line.
Closed — the outcome, kept so they can be told from the ones nobody wrote down
| What was deferred | How it ended |
|---|---|
| A wave's failure policy | carried in 2026-08-06, with its second occasion |
A pack mode for the inventory script | deleted 2026-08-06, by the owner's call |
| A nested layout — the machinery under one root directory | carried in 0.2.0, by the owner's call |
| Project-local skills must survive an upgrade that ships a same-named stock skill | reopened 2026-08-05, taken into the release |
| The contradiction stop has no form | shipped in 0.2.14 |
| The publish gate is tested against what its author expects bash to do | built 2026-09-24, the day its trigger fired: scripts/test-gate-vs-bash.sh, bash as the oracle, closed four ways |
| A lens round costs more than a session holds | adopted 2026-09-17: the re-round reads the repair's diff (lenses → Running them); its token cost stays open under The lens round has no form |
Bodies removed 2026-09-10. Each of these ended in a decision that shipped and left no open question, so the reasoning lives in the release that carried it and the row here says which. Two struck entries above are NOT in this table: the diagrams one because self-maintenance cites it by name, and the day-one cut because it still hides open work under a struck heading.