feat(comment): AI spam checking + real nested replies

[WebTigers]

AI spam checking

Tiger_Comment_Spam is the registry COMMENTS.md §6 designed, with the in-platform agent as its first checker. An Akismet-style module registers into the same seam.

• Shown only when an agent is connected — and the settings service refuses to store an enabled flag without one. A stored 1 with no agent looks like a working filter and silently isn't.
• No agent → silent no-op + one Tiger_Log line, comment passes through unchecked.
• isConnected(), not isAvailable() — the latter also asks whether the current user may chat, meaningless for a background check running for an anonymous commenter.
• A verdict may only TIGHTEN. spam bins it; ham / unknown / timeout / no-agent / broken-checker all leave the moderation posture alone. Nothing a checker says can publish something that wasn't going to be published — that's what makes it safe to hand attacker-controlled text to a model.
• Prompt injection is the live risk: the body is delimited and framed as data, only the two literal answers are accepted, anything else is unknown (fail-open). Worst case an injection wins the treatment the comment would have had with no checker at all.
• The poster is never told they were classified — a binned comment gets the same "awaiting moderation" reply, or a spammer iterates until it passes.
• One-shot complete(), not the agent Loop — a classification needs no tools, no ReAct steps, no transcript, and must not be able to do anything.

Cost/latency is real and acknowledged in the doc: an LLM round-trip per comment with a body, on the org's BYO key. Opt-in, skipped for star-only ratings, fails open on timeout.

Nested replies

parent_id was accepted by the API but unreachable — no reply affordance, and the thread rendered flat. Now:

• Default depth 3 (tiger.comment.threading, per-subject override).
• The list payload publishes the limit, so the Reply button disappears at the limit rather than failing on submit.
• The client orders the flat list into a tree, appending orphans rather than dropping somebody's words when a parent is removed mid-thread.
• Indent still caps at 3 so a deep thread doesn't walk off a phone.

Tests

36 new. One harness note worth flagging: several deliberately exercise fail-soft branches that log why they degraded, which trips beStrictAboutOutputDuringTests and reports a passing test as risky. They set the null log writer, so a covered branch doesn't have to choose between being tested and being quiet.

Full suite: OK (2183 tests, 21627 assertions).