fix: Resolve channel default_agent by name at message-time instead of caching UUID at startup

[psumo]

Problem

When a channel's default_agent is resolved at startup, the UUID gets cached in the AgentRouter. Kill + respawn assigns a new UUID, but the router still holds the old one. Every message after that fails with Agent not found: <old-uuid> until you restart the whole process.

I kept hitting this while iterating on agent manifests:

1. Edit manifest (tweak system prompt, change model, etc)
2. agent kill foo + agent spawn foo
3. Agent comes back with a new UUID
4. Telegram messages all break — router is still pointing at the dead UUID
5. Have to restart the entire container to fix it

Fix

Lazy re-resolution fallback in the bridge dispatch path. When send_message() gets an "Agent not found" error:

1. Look up the channel's configured default agent name (not UUID)
2. Call find_agent_by_name() to get the current UUID
3. Update the router cache
4. Retry the message

The cached UUID is still checked first so there's no overhead in the normal case. Re-resolution only kicks in when the send actually fails.

Changes

crates/openfang-channels/src/router.rs (+24 lines)

• channel_default_names: DashMap<String, String> field — stores the configured agent name next to the cached UUID
• set_channel_default_with_name() — stores both name and ID
• channel_default_name() — getter for the stored name
• update_channel_default() — swap out just the cached UUID after re-resolution

crates/openfang-channels/src/bridge.rs (+117 lines)

• try_reresolution() — checks for "Agent not found", looks up name from router, calls find_agent_by_name(), updates cache, returns new ID
• dispatch_message() error path — on "Agent not found", tries re-resolution + one retry
• dispatch_with_blocks() — same logic for multimodal messages
• blocks → blocks.clone() so blocks can be reused on retry

crates/openfang-api/src/channel_bridge.rs (1 line)

• set_channel_default() → set_channel_default_with_name() at startup so the name is preserved

Backwards compat

• set_channel_default() is untouched — only the startup path uses the new _with_name variant
• No new trait methods — uses existing find_agent_by_name()
• No config or API changes
• New fields on AgentRouter are private
• Existing tests pass as-is

Testing

1. Configure a channel with default_agent = "foo"
2. Send a message — works fine
3. agent kill foo → agent spawn foo
4. Send another message
5. Before: Agent error: Agent not found: ...
6. After: message goes through, logs show re-resolution happened
7. Subsequent messages hit the updated cache with no re-resolution

Edge cases:

• Agent killed but not respawned → falls through to the original error
• Non-"Agent not found" errors (model API failures etc) → no re-resolution triggered
• Multiple channels with different defaults → each re-resolves independently
• Multimodal messages → retry works with cloned blocks