feat(inbox): show unread mentions and add --mentions filter

[roryist]

When you run tdc inbox, you could not tell which unread threads had a fresh mention of you. The web app shows an @ badge for those. The CLI showed nothing.

The CLI was already fetching that information. The unread endpoint returns a mention flag for every unread thread, and the CLI reduced that response to a list of ids and threw the flag away. The directMentions field on a thread is not a substitute, because it only says who the first post mentioned.

Now every inbox thread carries hasUnreadMention in JSON output, there is a --mentions flag to show only those, and the human view marks them with a yellow @ and sorts them to the top of each channel. Channel thread listings get the same two fields from the same helper, so tdc channel threads --json and tdc inbox --json agree on shape.

Four tests cover the flag, the filter, the sort and the marker. Tests, type-check, lint and the skill-sync check pass. I ran the built binary against a real inbox and it returned exactly the threads the server flags.

One thing to know: this means "unread mentions", the same as the web badge. A mention you have already read past will not show. An "ever mentioned" filter needs the backend to expose a field it currently hides.

(Written by Rory's agent)

🤖 Generated with Claude Code

https://claude.ai/code/session_01XYtXq3VLuv34zPKicow2jg

[roryist]

@doistbot /review

[doistbot]

This PR adds unread-mention awareness to the inbox: every thread now carries hasUnreadMention in JSON output, a --mentions flag filters to threads with fresh mentions, and the human view marks them with a yellow @ and sorts them to the top of each channel. The shared unreadFlags helper keeps tdc inbox --json and tdc channel threads --json consistent, and the tests cover the flag, filter, sort, and marker.

Few things worth tightening:

• The unread lookup keeps full UnreadThread objects for the rest of the command when consumers only need existence and the directMention boolean. Since getUnread() is workspace-wide, storing a compact Map<string, boolean> (or similar flag value) would avoid retaining fields like channelId and objIndex for large unread sets.

I also included a few optional follow-up notes in the details below.

Optional follow-up notes (3)

• src/lib/threads.ts:6: CODEBASE.md still references the removed fetchUnreadThreadIds helper (lines 162 and 179) after this rename to fetchUnreadThreads. Update those references so the repo map matches the exported function.
• src/commands/channel/threads.ts:127: tdc channel threads --json now emits hasUnreadMention, but no test covers it. The PR explicitly advertises shape parity with tdc inbox --json, yet src/commands/channel/threads.test.ts still only asserts isUnread/url. Extend the existing --json emits isUnread and url without --full test (or add one) to assert hasUnreadMention: true for an unread entry with directMention: true.
• src/commands/inbox.test.ts:275: This sort test has exactly one thread per tier (mention / plain unread / read), so it only proves the tier ordering. A comparator regression that dropped the || sortByDate(a, b) tie-break would still pass, because each tier contains a single item and the fixture's original order already matches the tier order. Add a second thread to one tier (e.g. another plain-unread thread with a different posted value) and assert newest-first within that tier.

_{Share Feedback • Review Logs}