Loop-to-loop discussion is invisible on the board: 'direct' replies mirror as records and hide in a rollup that forgets it was opened

[scgopi]

Loop-to-loop replies are mirrored as record posts, and the rail's board hides every one of them behind a one-line rollup that forgets it was opened. When two loops actually discuss something — a root-cause correction, an answer to a review question — the supervisor watching the board sees nothing.

What happened

Two loops (TripleAgentSpawn, IssueTriage) worked through the diagnosis that became #272 entirely over graphcode node send. Five posts landed on the board; one was visible.

Post	Topic	Kind	Content	Visible in the rail
#69	direct	record	Reported the codex-pane bug + a test-coverage question	❌ rollup
#70	issue-272	note	"Filed #272 and CLAIMING it"	✅ board list
#71	direct	record	Heads-up to the Release loop about 0.1.61-beta1	❌ rollup
#72	direct	record	Root-cause correction: truncation is not the mechanism, zmx run sets .command = null	❌ rollup
#73	direct	record	Correction accepted + verified, answered the test-coverage question	❌ rollup

Store-wide on this project right now: 50 notes, 23 records.

The two posts carrying the actual technical substance (#72, #73) are the two least visible.

Why

graphcode/Sources/Features/LoopWorkspace/ArtifactorySection.swift

Line	Behavior
:19 notes(in:)	Main list filters to kind == .note
:26 records(in:)	Direct messages and handoffs split off — "receipts for deliveries that already happened, not something written to be read here"
:34 unreadNoteCount	Records excluded from the badge — "a badge counting mail nobody is being shown is a badge that cannot be cleared"
:62	Records-expanded state is local and unpersisted
:238-269	recordsRollup renders "N message records", expanding to the last 8; older ones have no affordance at all

Each decision is defensible on its own. Together they mean a channel loops genuinely use for discussion is one a supervisor never reads — which the section's own doc comment names as the failure mode it exists to prevent:

a coordination channel a supervisor never sees is the failure mode, not a missing convenience

The real problem

The direct topic does double duty. It carries both:

• receipts — "handoff delivered", "message staged" — which genuinely are noise, and
• discussion — corrections, answers, review findings — which is exactly what the board is for.

kind == .record is applied on delivery mechanism, not on content, so the two are indistinguishable downstream.

Possible directions

1. Persist the expanded state. Smallest fix. :62's local @State becomes an AppStorage key like the section's own fold. Doesn't fix the badge or the >8 cutoff.
2. Count records in the badge once they can be shown and cleared, removing the reason :34 excludes them.
3. Separate receipt from discussion. A delivery receipt is a record; a message a loop wrote is a note. This is the fix that matches the stated intent — the mirrored body is already the full message text, so the information to make the split is present at mirror time.
4. Drop the last-8 cutoff in the rollup, or give "N earlier" a way to open.

(3) plus (2) is my read of what the section's own reasoning is asking for, but the call is a design one.

Repro

1. Two loops in a project; from one, graphcode node send <other-id> <a substantive message>.
2. Have the target reply the same way.
3. Open either loop's rail. The exchange is not in the board list; it is inside a collapsed "2 message records" line, and there is no unread badge for it.
4. Expand it, navigate away, come back — it is collapsed again.

Found while verifying #271 and diagnosing #272.