feat: sentinel review — the archive owner's verdict, recorded on the file

[khaneight]

Stacked on #70. The permission system: approve, reject, request changes, comment.

$ sentinel review
Waiting on you — 1 item(s)

  [trait] distrusts-abstraction
      Treats an abstraction as unearned until something concrete has failed.
      unconfirmed reading of the author · persona/distrusts-abstraction.md

$ sentinel review distrusts-abstraction --reject \
    --note "too strong — I distrust *unearned* abstraction, not abstraction"
rejected — persona/distrusts-abstraction.md (2026-08-25)
  status is now `rejected`

No skill invokes it. It joins export in NOT_FOR_AGENTS: an agent that
can approve its own work has a permission system in name only.

Where verdicts live, and why they append

In the document's own frontmatter. A verdict recorded in meta/ comes apart
from the thing it was about the first time a file is renamed by hand.

They append rather than overwrite, because the history is the point — changes
requested → note → approved three weeks later is the useful record, and a
single mutable field keeps only the last word. The operative verdict is the
latest that decided something; comment deliberately isn't one, so a remark
left on approved work does not un-approve it.

Three refusals

• A verdict signed by nobody. With no --by, SENTINEL_REVIEWER or USER
  it stops rather than defaulting. A record attributed to a placeholder looks
  exactly like somebody agreeing, which is the one outcome this whole mechanism
  exists to prevent.
• An ambiguous name. A trait and an article can share a slug. Recording a
  verdict on the wrong document is worse than recording none, so it names both
  and stops.
• A partial view. The unreadable file could be the one the name really
  meant — or the one that made it ambiguous.

The append is textual

Like mv's citation rewriting, and for the same reason: the block is a file the
user also opens by hand, and a serde round trip would reorder its keys and strip
its comments to add four lines. Tested against a comment above title:, an
inline tags: [a, b], a key after the list, and an indented review: nested
under something else.

The entry itself is serialised, so a note containing colons, quotes and
newlines is YAML's problem rather than this function's. A review: already
written as a populated inline list is refused with instructions rather than
mangled.

Traits carry their standing twice

status: is what a reader sees at the top of the file; review: is the history
behind it. Deriving one silently from the other would mean a file that reads
proposed while the archive treats it as affirmed — so both are written
together and verdict-disagrees-with-status reports them drifting. Articles get
no such treatment: their status is maturity, and stable means finished,
not signed.

CLAUDE.md's budget: 12,000 → 14,000

Raised after compressing, not instead of it. The repository file absorbed a
second document schema, a permission system and a ladder rung while growing 325
bytes net: mv/rm reasoning and the case-sensitivity recipe moved to
docs/design-notes.md, the command reference lost what --help already says,
and "derived sets" and "one source of truth" turned out to be one invariant
written twice. The reasoning is recorded on the constant, with an instruction to
compress first and say what came out before raising it again.

484 tests (+25), clippy and fmt clean.

🤖 Generated with Claude Code