fix(feedback): make comment delivery exactly-once across channels

[benvinegar]

Problem

The server's agentSeq cursor ("highest comment seq already delivered to the agent") was advanced by every delivery channel but read by none. Each client kept a private cursor defaulting to 0:

• the CLI (sideshow wait) in a per-process state file
• the stdio MCP server in a process-local variable
• the /mcp transport expected the model to pass afterSeq back, defaulting to 0

So mixing channels — exactly the combo the docs recommend (piggyback on writes + a background sideshow wait) — replayed comments the agent had already received. Found in the wild: a background sideshow wait armed after an MCP wait_for_feedback immediately re-delivered the same comment.

Fix

• server/app.ts: an author=user session read with no explicit after now resumes from the session's agentSeq — "where the agent left off" lives server-side and is shared by every channel. An explicit after/afterSeq remains as a deliberate replay (and can't regress the cursor; markAgentSeen is monotonic in both stores).
• bin/sideshow.js, mcp/server.ts, server/mcpHttp.ts: the three client-side cursors are deleted; clients omit after and let the server resume.
• Viewer behavior unchanged: unfiltered reads never touch the cursor.
• Docs updated (guide/, MCP tool descriptions, AGENTS.md invariant, CHANGELOG).

Tests

• New: cursor-less author=user reads deliver once and not again across fresh processes; piggyback delivery advances the cursor seen by waits; viewer reads stay unconsumed.
• Updated: the author/after filter test passes after=0 explicitly, since a bare author=user read now (correctly) consumes the cursor.
• npm test (51/51), npm run typecheck, npm run lint, npm run format:check all pass; also verified live against a throwaway server (deliver once → fresh process gets nothing → piggybacked comment not re-delivered → --after 0 replays).

🤖 Generated with Claude Code