fix(delivery): preserve exact identity pairs in check-inbox

[suiren]

Summary

Derive check-inbox.sh's delivery targets from identities.sh's exact
(team, agent) rows instead of whoami.sh's flattened human-readable
output, so the Stop hook never polls — or marks read — a (team, agent)
pair that was never registered.

Problem

identities.sh returns the registered pairs as team<TAB>agent rows:

alpha	alice
beta	bob

whoami.sh flattens those rows into independent lists for its
human-readable output:

multiple=true agents=alice,bob teams=alpha,beta

check-inbox.sh parsed that output, taking the first agent and
every team. With multiple identities registered for the same
(project, type) it therefore polled the cross product — here
(alpha, alice) and (beta, alice) — although (beta, alice) was
never registered. Any message row matching such a phantom pair is
displayed to the wrong session and marked read, so its actual
addressee never sees it.

Change

Consume identities.sh's TSV directly:

• Keep the existing first-agent policy, but accumulate only that
  agent's actual team rows — no cross-product pair is ever queried or
  marked read.
• Keep the project-root resolution the whoami path performed
  internally: agmsg_resolve_project runs before the lookup, since
  identities.sh itself is an exact registry lookup by design (its
  other callers depend on that). Without it, a session started from a
  nested subdirectory or a sibling worktree would silently stop
  receiving messages.
• Empty or malformed identity output delivers nothing, matching the
  previous not_joined/suggest behavior.

Unchanged: the cooldown marker stays keyed by the selected agent, the
per-pair actas exclusivity check still runs inside the team loop, and
the #653 loop-failure capture (CLAIM_RC) is untouched.

The empty-array expansion in the dedup loop is guarded with the
${arr[@]+...} idiom already used in resolve-project.sh, since the
script runs under set -u and macOS CI runs bash 3.2.

Test plan

Two regression tests in tests/test_inbox.bats:

1. With (alpha, alice) and (beta, bob) registered for the same
   project and type, only (alpha, alice) is displayed and marked
   read; both cross pairs and the second agent's row stay unread.
   Reverting the fix turns it red: the old parser polls
   (beta, alice) and falsely marks it read.
2. With a registration at the project root, an invocation from a
   nested subdirectory still resolves to the root and delivers —
   guarding the resolution step, which no existing test exercised at
   the check-inbox call site.

Honest status: the suite is expected to run in this repo's PR CI; it
was not executed locally on our side (Windows dev machine, as with
#613).

Note for integration/remote

The same flatten-and-cross-product block exists unchanged on
integration/remote (the storage rework there did not touch the
identity derivation), so this fix will want a small port when that
branch lands on main. Both regression tests should carry over as-is.

---

🤖 Generated with Claude Code

[fujibee]

@suiren

Thank you for this, and I'm sorry it sat this long without a word from us. Merged.

We reproduced it on main first: with alpha/alice and beta/bob both registered for one (project, type), a message sent to beta's bob is marked read by the Stop hook, because the targets are built as the first agent crossed with every team — a pair that was never registered.

The approach is right. Reading identities.sh's (team, agent) rows instead of whoami.sh's flattened output is what watch.sh already does per pair (#62), so both delivery paths now derive targets the same way. It also skips whoami.sh's additional scan of every team on disk, so the hook does slightly less work rather than more.

We re-ran CI before merging rather than trusting the existing green — it was against a base from six days and forty-four commits back, so it was real but no longer current.

This goes out in the next release.