Record the deletion gap, and bring the counts up to date

[R204570]

Both documents predate #23. Updated ahead of tagging the first release.

AUDIT.md — F14 and a new §10

Deletion was not found by auditing behaviour. It was found by trying to clear
the wrong astro harvest from §9 and discovering there was no way to.

kb_store had delete() on both backends the whole time — tested, working,
and unreachable. No route, no UI, no CLI flag, no tool. Every wrong harvest was
therefore permanent.

That gap reads worse against the rest of this audit than it would in isolation:

An audit whose central finding is "the store confidently holds wrong things"
cannot also be one where wrong things are permanent. Every honesty signal
exists to tell a caller a copy is bad. If the answer to "so remove it" is
you can't, the signal is a label on a problem nobody can act on.

§7.3 updated. The store still holds the pre-fix corpus and that is still the
one action outstanding — but it is now a supported operation with a command
against it rather than a wish, and it is the owner's to take, since the harvests
are theirs.

PROPOSAL.md — §4.8

The same point as architecture: a store needs a full lifecycle, not an append
path. Three surfaces, because the person who took a harvest is the one who
decides it was a mistake — and deliberately not the model by default, since
by §4.1's own argument the caller most likely to want a delete is the one that
has just mis-resolved a name.

Also

• Test counts 346 → 375 in both.
• F14 added to the status table in the audit and the findings table in the
  proposal.

Verified

• 375 passed, 22 skipped across both backends. No code changed —
  documents only.
• All four mermaid diagrams rendered in Chromium via mmdc.