refactor(daily-catalog): make UserAgentCatalogGAgent pure set-membership; projector consumes SkillRunner committed events directly

[eanzhao]

Architectural follow-up surfaced in docs/audit-scorecard/2026-04-27-daily-pipeline-architecture-review.md §B1 + §B6. This is the structural fix behind #440; #440's own "Suggested fix direction" only lists patches (coalesce / defer / watermark) that leave the underlying coupling.

Symptom (recap)

UserAgentCatalogGAgent is a well-known single actor that wears two hats:

• set-membership owner (which agents exist) — long-lived fact owner ✅
• per-entry execution-status hub — receives UserAgentCatalogExecutionUpdateCommand from every SkillRunnerGAgent after every run ❌

The early-return guard in HandleExecutionUpdateAsync (agents/Aevatar.GAgents.ChannelRuntime/UserAgentCatalogGAgent.cs:91-114):

```csharp
if (State.Entries.All(x => !string.Equals(x.AgentId, command.AgentId, StringComparison.Ordinal))) {
Logger.LogWarning("Cannot update execution state for missing user agent catalog entry: {AgentId}", command.AgentId);
return;
}
```

silently drops the trigger-side execution update if it arrives before the init-side upsert has been applied to state. That is the architectural root cause of #440 "first-run not reflected in /agent-status".

Architectural violations

• CLAUDE.md "单线程 actor 不做热点共享服务" — every agent's every execution funnels through one well-known actor.
• CLAUDE.md "Actor 即业务实体" — catalog is an aggregate view, not a per-execution coordinator.
• CLAUDE.md "actor 即业务实体... 数据与方法同住" — SkillRunnerGAgent shouldn't know catalog's command schema; it does today (see #B6 in the audit doc).

Proposed direction

1. Catalog actor handles only UserAgentCatalogUpsertCommand / UserAgentCatalogTombstoneCommand. Remove UserAgentCatalogExecutionUpdateCommand from its message set.
2. SkillRunnerGAgent (or its split successors per #refactor-split-skill-runner) commits SkillRunnerExecutionCompletedEvent / SkillRunnerExecutionFailedEvent only.
3. UserAgentCatalogProjector subscribes to both event streams (catalog upsert/tombstone + SkillRunner execution events), keyed by agent_id, and writes to UserAgentCatalogDocument with covering-write semantics.
4. The early-return guard goes away — projector covers by primary key, late events still land correctly.
5. SkillRunnerGAgent.UpdateRegistryExecutionAsync and the UserAgentCatalogExecutionUpdateCommand proto / handler are deleted (net code removal).

Acceptance

• After /daily ... with run_immediately, /agent-status reflects Last run / Next run within projection lag SLO. (Same as bug(agent-status): SkillRunner first-run completion does not reflect in Last run / Next run #440 acceptance.)
• No "Cannot update execution state for missing user agent catalog entry" warnings under any timing.
• SkillRunnerGAgent no longer references any UserAgentCatalogXxx command type.
• Test: spawn N concurrent /daily creates, all of them reflect committed status in catalog readmodel without races.
• Net diff is negative (deletes the command path).

Affected files

• `agents/Aevatar.GAgents.ChannelRuntime/UserAgentCatalogGAgent.cs`
• `agents/Aevatar.GAgents.ChannelRuntime/SkillRunnerGAgent.cs` (UpsertRegistryAsync / UpdateRegistryExecutionAsync sites)
• `agents/Aevatar.GAgents.ChannelRuntime/UserAgentCatalogProjector.cs`
• `agents/Aevatar.GAgents.ChannelRuntime/AgentBuilderTool.cs` (UserAgentCatalogUpsertCommand call sites)
• `agents/Aevatar.GAgents.ChannelRuntime/channel_runtime_messages.proto` (UserAgentCatalogExecutionUpdateCommand removal)

Related

• bug(agent-status): SkillRunner first-run completion does not reflect in Last run / Next run #440 — bug-level symptom; ship the patch there for now, this is the structural fix.
• bug(daily): GitHub username binding shared across all Lark users of one bot — last writer wins #436 / [Bug] /daily binding causes cross-user data leakage #437 / bug(daily): SkillRunner masks GitHub tool failures as silent "no activity" successful runs #439 — separate axes of the same /daily flow.