[omartelo]

Summary

A survey of the session title across all eight providers found Claude Code and Codex cards no longer titled, Cursor CLI titleable, and Crush not reliably so. This PR fixes the lich side; the Codex and oh-my-pi fixes are in lichdotdev/lich-plugin.

• Claude Code: a new session is spawned without --name. The flag marks the name as the user's (nameSource: "user"), and a session named that way never writes ai-title (measured on 2.1.306; no transcript on disk from 2.1.245 on holds both). Unnamed, Claude Code names itself <dir>-<2 chars>, distinct within one checkout (measured with three sessions in one directory), but writes that name only to ~/.claude/sessions/<pid>.json. So relay.RosterNameOf returns no second name for a Claude session with nothing on record, instead of a derived one its own roster never shows. A fork is still named, and a /rename is still honoured.
• terminal.Start drops its name argument, which no caller used any more; native/lichclient moves with it.
• open_session says "It answers to X." when there is no second name (new OpenedSessionLabelOnly, all seven locales).
• Cursor CLI: the TUI files a generated title in the meta.json beside the chat's store.db; lich reads it on the stop report, alongside Kiro (title_kiro.go becomes title_filed.go).
• Crush: the gap is written down in docs/ceilings.md and docs/hooks/session-title.md; the README provider table gains a "card names itself after the conversation" row in every language.

Tests changed because the contract did: Start no longer takes a caller's name, a new Claude session is unnamed, and a Claude peer with nothing recorded has no roster name.

Test plan

• gofmt -l ., go vet ./..., go test ./... (backend)
• native/: gofmt -l ., go vet ./..., go test ./...
• cross-compile loop for linux, darwin, windows
• Claude Code 2.1.306 measured with and without --name: ai-title only without, roster names distinct without
• Cursor CLI 2026.10.16 measured: meta.json title within 4 s of the first prompt in the TUI