feat: the corpus is read before the clone writes, and grounding is not authorship

[khaneight]

Stacked on #81. Builds the plan from #82.

① fresh corpus
Next: learn
  1 document(s) registered `origin: authored` that no persona trait has been read from

② corpus read, not yet answered
… 1 trait(s) are waiting on your verdict. The archive has read the corpus;
  nothing below this writes anything until you say whether it read you right.
  sentinel review

③ after affirming
Next: compile

learn moves above compile

It was below on a cost argument — compiling a document is the close reading
that makes mining it cheap. True, and it answers the wrong question. The order
follows what the tool is for.

next stops for a verdict

With traits proposed and none affirmed it returns none and points at sentinel review. My earlier objection was that this deadlocks a new archive; it doesn't,
it is the sequence.

Only when there is something to answer — an archive that never built a persona
is not waiting on anybody and is not blocked. Tested both ways.

It renders … rather than ✓, because three different things reach that branch
and only one is finished.

Every article records the voice it was written through

unvoiced-article — a warning, and only once the archive has an affirmed
trait. An article written before the persona existed is unfinished, not
malformed, and warning in an archive with no persona is noise about work nobody
could have done.

/sentinel-improve says the repair is to cite what an article genuinely
exhibits and to report the ones that exhibit none — a flat summary of someone
else's paper often won't. Attaching plausible traits to clear a warning is
inventing the authority the article was written on.

compiles → grounds, and it is no longer a way in

The old model treated sources: as authorship — document → article, as
though the document wrote it. It didn't; the clone did, through the persona. A
citation reaches back past whatever produced the text, so it is marked
primary: false, the hover walk ignores it, and the page draws it dashed.

Authorship radiates. Three rings, one path out, no fourth ring needed.

Two mistakes of mine

The proposal in #82 argued from an edge count on a demo I had built with the
current ladder — measuring the ladder, not the world. You caught it.

And I reverted this rule once with a stray git checkout inside a compound
command. The rule-coverage guard caught it by reporting the rule as never
firing, which is the second time that guard has earned its place.

Coverage

tests/onboarding.rs asserts the whole sequence, since this changes what a user
sees in order — including that the waiting state does not print a tick.

543 tests (+3), clippy and fmt clean. CLAUDE.md compressed under budget again.

🤖 Generated with Claude Code