docs(attr-child-clock): add reference findings and the feasibility fact

ATTRIBUTION-child-clock.md

@@ -89,3 +89,76 @@ Say which category your change is in.
    does not.
 4. Propose a mechanism; prefer architecture over a guard. State whether it removes the bug class.
 5. Mark every claim VERIFIED or UNVERIFIED. Cite file:line.
+
+---
+
+# Reference findings
+
+Five independent implementations were reviewed. Projects are deliberately not named here; describe
+any conclusion in repo-native terms only. Every mechanism below was read at source and verified.
+
+## Converged — all five agree
+
+1. **A child's activity never touches the parent row's lifecycle state, label, or timestamp.**
+2. **Attribution happens at ingestion (the producer boundary), never as a reader-side filter.**
+3. **Exactly one writer of a row's status/recency.** None of the five keeps two writers apart with
+   publication filters. They make the second writer *impossible*.
+4. **Child -> parent rollups do exist** — but they roll up *status* or a *count*, never a label and
+   never a timestamp, and each is explicit and documented.
+5. **The row label is a stored title derived from the user's own prompt** — never a backward scan
+   for the newest assistant message.
+
+## Closest architectural match (same platform: Electron, host-service process + renderer sidebar
+## that nests parent and subagent rows)
+
+- Subagent hook events are routed to a **separate roster at ingestion, before any store write**.
+  The ingestion comment reads: *"Subagent activity is not the terminal's lifecycle: no chime, no
+  status change, no session id capture."*
+- The store method recording subagent events is documented *"Never touches the parent binding's
+  lifecycle state."*
+- The renderer **derives** status on every read inside a `useMemo` and **stores nothing**.
+- The state-entry stamp is preserved server-side as
+  `prior !== undefined && !sessionChanged ? prior.startedAt : occurredAt`.
+  **There is no done-state special case.** That is the shape of Orca's *main* canonical writer
+  (`server-ingest-structured.ts:65`), not the renderer bridge's.
+
+## How the references make a second writer impossible
+
+- One reference's generic update API **structurally omits** the status/recency fields, so the type
+  itself forbids a second writer; a lint rule confines cache writes to owner modules.
+- One consolidated *"ten independent writers across six slices"* into a single registry writer and
+  recorded the decision as an architecture decision record.
+- One gates client updates behind a monotonic sequence check and replaces **whole rows**, so a
+  client cannot mutate an individual field.
+
+## Where the references DIVERGE — do not claim precedent here
+
+The child clock (defect C) has **no single answer**:
+- one shows **no timestamp at all** on the desktop row, carrying liveness via status + derived counts;
+- one shows a **live-ticking elapsed duration** that freezes at completion;
+- one carries a true **advancing per-child progress field**.
+
+All three do quarantine the spawn stamp from display. So C has several legitimate answers — choose
+one with stated reasons; you may not justify it by precedent alone.
+
+## Material deviation to disclose (repo-native terms, for the eventual PR body)
+
+> Orca keeps two real writers of one status row apart with publication filters. Every reference
+> implementation examined removes the possibility of a second writer rather than suppressing its
+> traffic.
+
+## Feasibility fact established in Orca
+
+The producer already has the attribution and discards it:
+- `src/main/claude/claude-structured-item-translation.ts:57` already parses
+  `parentToolUseId: claudeText(frame.parent_tool_use_id)`.
+- `src/main/claude/claude-structured-journal-translation.ts:143-144` uses it only for
+  `observeChildActivity(...)`, then drops it; the item is journaled with no producer trace.
+- The journal schemas carry **no producer field at all** — zero hits for
+  `parentToolUseId|agentId|parentAgentId|producer|subagentId` in
+  `src/shared/agent-session-journal-schemas.ts` and `src/shared/agent-session-journal-types.ts`.
+- A canonical subagent id already exists next door:
+  `src/main/claude/claude-subagent-roster.ts:121` (`this.ids.canonical(parentToolUseId)`).
+
+So attributing at the producer means **persisting a value we already compute and throw away three
+lines later** — not plumbing a new one end to end.