Persistent observation log — foundation for the contextualizer learning model

[smorchj]

Context

PR #78 landed the live session watcher (#77). It tails Claude Code JSONLs in real time and feeds the activity store for the live node graph. But the watcher is ephemeral — events flow through for the 8-second pulse lifetime and are then lost.

The learning model (repetition + emotion signals) needs to compute stats across sessions: "this file has been read 14 times across the last 10 sessions", "this folder is consistently missed by routing", "Claude hit max_turns here three times this week". None of that is possible without persistent history.

Scope

Add an append-only observation log stored under `.klonode/observations.jsonl` per project. Every tool_use event the server watcher extracts gets a line in this file.

Schema

```json
{
"at": "2026-04-15T21:11:34.455Z",
"sessionId": "120e914c-6c31-49a5-99a7-0ec67b527a13",
"cwd": "C:\Users\smorc\Desktop\KlonodeV2",
"tool": "Read",
"kind": "read",
"path": "packages/core/src/analyzer/scanner.ts",
"isSidechain": false,
"uuid": "2ecbe75a-...",
"toolUseId": "toolu_01Jp..."
}
```

Paths normalized to forward slashes and relative to the session's cwd (same normalization the activity store does, so downstream learning can join directly against graph node paths).

Server

• New module `packages/ui/src/lib/server/observation-log.ts` — append, read by range/session/path, tail for new-only reads.
• The session watcher writes each emitted event through this log before (or alongside) pushing it over SSE.
• Dedup by `uuid + toolUseId` so restarts and reconnects don't double-record.
• Bounded — rotate at say 50 MB per project, gzip older chunks.

CLI

• `klonode observations status` — counts per tool, per session, per path prefix.
• `klonode observations purge` — drop everything, start fresh.

Out of scope

• Doing anything with the data yet. That's the next issues (learning model repetition + emotion).
• Machine-wide observation log. This one is per-project so users can check in the graph but leave observations out (or vice versa).

Dependencies

Depends on nothing — builds directly on the server session watcher from #78.

Part of the #77 pivot roadmap.