Retire the documentation's pre-implementation status

[wolpert]

Every status paragraph in the corpus was written before a port existed and said so. The Java port
now reaches every conformance level and declares all of them in conformance/declarations/java.json,
so a reader was told that no implementation exists by the same paragraph that went on to describe
one. A review of the whole corpus then found thirteen further statements the tree contradicts, plus
three register and justification defects the style check does not read.

Commit 1: the status paragraphs

The status lines of 00-overview.md, 05-glossary.md, 10-specification.md,
20-topology-format.md, 30-conformance.md, 40-java-binding.md, 90-open-questions.md, and
99-roadmap.md now name ports/java/ as the one implementation. Each keeps the part that is still
true: a second port has not been started, nothing is published, and the requirements still state
what a conforming implementation does rather than what one did.

README.md, docs/README.md, and CLAUDE.md carry the same repair. README.md gains the JMH
benchmark caveat that ports/java/README.md and 40-java-binding.md already state.

style.md and CONTRIBUTING.md carried the worse defect. Both closed their enforcement section
saying neither documentation check exists and that every rule is enforced at review until the first
Java implementation lands. Both checks have run in that port's check task since 157472a, over
every Markdown file in the repository, and style.md contradicted itself within one section by
opening with a table of two checks that run.

Two RFC 2119 leaks outside 10-specification.md are repaired with them, both reported by
verifyDocStyle. ADR 0070 argued an alternative in terms of MUST and MAY; it now says
requiring and permitting. ADR 0078 and style.md each quote a keyword as a keyword, so both are
marked as code spans.

Commit 2: ten claims the tree contradicts

Where	Claim	Contradicted by
CLAUDE.md:24	regenerating the suite is the only build in the repository	ports/java/ is a Gradle build, documented fifty lines below
CONTRIBUTING.md:3	the library "is being implemented"	the same stale register retired elsewhere
CONTRIBUTING.md:5	"it reaches the every conformance level"	a broken edit that survived
docs/README.md:156	"the Gradle projects", "what each split lets a consumer avoid"	adr/0081 left one artifact and one module
docs/README.md:161	"the build check that is to enforce it"	verifyUnsignedComparisons is registered and wired into check
99-roadmap.md:234	sharder-api, sharder-core, sharder-migrate, sharder-provider-file	adr/0081 records these names as unpublished and free
35-port-conventions.md:124	./gradlew :sharder-conformance:test	no such Gradle project; a second port copies this example
40-java-binding.md:1103	this binding "has not been written"	core/internal/hash/SipHash24.java
40-java-binding.md:1373	verifyDocLinks walks docs/, ports/, and README.md	the include list also carries CONTRIBUTING.md
90-open-questions.md:236	the probe counter and budget window, "and neither takes a lock"	RetryBudget.permitted and .account are synchronized

The roadmap one is the highest value: naming superseded artifact coordinates inside the one-way
doors section puts them where the document says nothing can move.

The retry budget one is the subtlest. The probe counter is an AtomicInteger, but RetryBudget
holds its window under one lock deliberately, because FAIL-031 counts to the millisecond and a
bucketed counter answers differently at a bucket boundary. ports/java/README.md and the class
javadoc both record that difference. The open question asks whether a requirement should forbid a
lock on these counters, so the port is evidence bearing on it, and the entry now states that rather
than contradicting it.

Commit 3: register and justification, and one stale open question

verifyDocStyle reads four mechanical rules and does not read register, justification, or
terminology, so a green check does not clear these.

35-port-conventions.md:41 and adr/0082:20 each asked what the repository wants from a port's
build in the second person, "build yourself, and exit non-zero where you failed". The Register
section admits second person only in the numbered steps of a procedure, and this is expository
prose in both. The ADR is not exempt: the Justification carve-out covers what a record may discuss,
not how it may sound. Both now ask whether the port builds and whether it exits non-zero. No second
person now remains in docs/, conformance/, or the root documents outside DESIGN_PROMPT.md and
the examples style.md quotes in order to forbid them.

conformance/generator/README.md argued that a hand-written vector file is worse than no vector at
all "because no correct implementation can pass it and every implementer assumes the defect is
theirs". The second clause is about the reader's assumptions, and the argument belongs in a
decision record rather than a directory's entry point. The fact it protects is already the sentence
before it, so one plain sentence remains.

OQ-11 asked whether the reference implementation checks the specification or only itself, and
named as its settling evidence a Java port written from the specification reaching core without
forcing a vector to move. That port has since reached every level, and it did force vectors to
move: adr/0084 and adr/0085 each record a place the reference and the specification disagreed.
The entry read as though none of it had happened. It now records the outcome and keeps the question
open on the release column's own terms, a second port being what shows the specification reads the
same way twice, rather than re-deciding its status.

Checks

• ./gradlew verifyDocLinks resolves every cross-reference across 99 documents.
• ./gradlew verifyDocStyle reports no findings, down from five.
• ./gradlew build green, including every other gate.
• verify_withdrawals.py, verify_declarations.py, and coverage.py --check pass.
• No generated file is touched; the change is Markdown only.

No decision record accompanies this, because nothing here turns on a judgement somebody could
reasonably make the other way. Every edit corrects a statement the tree already contradicted or a
rule style.md already states.

🤖 Generated with Claude Code