A placement record written by the action hook has no project or type, and arrange refuses it

[fujibee]

agmsg_self_name_on_action <team> <agent> [<project>] [<type>] takes project and
type as optional, and all three action callers omit them:

scripts/send.sh:32     agmsg_self_name_on_action "$TEAM" "$FROM"
scripts/inbox.sh:23    agmsg_self_name_on_action "$TEAM" "$AGENT"
scripts/history.sh:26  agmsg_self_name_on_action "$TEAM" "$AGENT"

The hook passes them straight through to the primitive, which writes the
placement record as <ref>\t<project>\t<type>. So a record written by the
action hook — the #1109 path, the one that exists for hand-started seats — has
two empty fields.

arrange.sh reads those fields and refuses:

IFS=$'\t' read -r ref project type < "$rec" || true
...
[ -n "$SOURCE_PROJECT" ] && [ -n "$SOURCE_TYPE" ] \
  || die "placement record for '$TEAM/$SOURCE' is missing project or type"

Reproduced end to end in an isolated skill dir, with the differential control:

source = a record with empty project/type
  -> arrange: placement record for 'T/a' is missing project or type

source = a record that has them
  -> gets past that line and fails later for an unrelated reason
     (the fixture has no roster entry)

Measured on one machine's live store: 12 of 36 placement records have an
empty project or type, including seats that are in daily use. Those seats cannot
be used as the source of an arrange.

The two ways out look different in what they claim:

• the callers pass the project and type they already have (send.sh and inbox.sh
  both know $PROJECT/$TYPE), so a record written on action carries the same
  fields a record written at spawn or SessionStart does; or
• arrange stops requiring fields that a legitimate writer does not supply,
  which means deciding what it actually needs them for.

The first is smaller and keeps one shape for the record. Filing rather than
fixing because it is not in the path of the release-gate work in flight, and
because the second option is a question about arrange, not about the hook.

Found while checking that the #1130 fix had reached this machine's seats.

[fujibee]

Folded into the inversion (#1152) rather than fixed where it is. The defect survives the change of author: whoever writes the record has to be given project and type, and today's callers simply do not pass them. Writing it correctly in the new path costs nothing; adding it afterwards is a second pass.

Worth carrying the distinction with it: a test that the record was created is not a test that it was created correctly. That gap is how this one outlived the change that introduced the record write.