fix: attribute recorded sessions to the working directory they ran in

[Diego Colombo (colombod)]

What this ships

The Context Intelligence server records the working directory each session ran in and stores it on the Session node, so it can be queried afterwards.

What it fixes

The hook has emitted working_dir as a top-level envelope field since the client-side change in August, and post_events has been persisting it verbatim with every event — but the drainer discarded it when parsing the queue line, so it never reached the graph.

Note on scope: workspace is already the slugified working directory, so grouping sessions by project works today. What was missing is the exact path — slugging is lossy (/home/user/my-app and /home/user/my/app are indistinguishable once slugged), and the only inverse available is a best-effort approximation. This recovers the real path.

Behavior

• working_dir is read off the durable queue line, not bound to the in-memory worker. That is what makes it survive a restart: a drainer respawned by crash recovery or dead-letter replay never sees the original HTTP request, but it replays the same bytes. No changes to either recovery path were needed.
• It is threaded as an explicit keyword to the Session-node write rather than injected into data — data is stored verbatim as a blob on every Event node, so injecting it there would duplicate the path per event.
• Populate-if-missing: written when the node has no value, never overwritten once set. Enforced twice — in ensure_session_node, and again at the Neo4j MERGE via coalesce(n.working_dir, row.working_dir), so the rule also holds for concurrent writers and replayed batches, not just for writers that went through this process's cache.
• Because the backfill also applies to nodes that already exist, re-importing a session recorded before this change fills the value in.
• Sessions that never report a working directory leave the property unset. None means "not reported", which is not the same as the empty path; whitespace-only values are rejected at ingest.
• The value lives on the Session node only. Every Event is one HAS_EVENT hop away, so there is no per-event copy on the highest-volume write path.

Tests

Full suite: 1967 passed, 4 skipped — including all 79 live-Neo4j tests against a real container. ruff format/check clean; pyright error count unchanged from main.

New coverage:

• the envelope field surviving to the durable queue line (and staying out of data)
• a fresh worker draining a persisted line — the crash-recovery shape
• backfill of a Session node that predates working_dir
• non-overwrite when an event reports a different folder
• live-Neo4j gates proving the coalesce holds under a second writer, and that a row carrying no working_dir does not null an existing value
• _parse_line normalization (absent and "" both yield None)

Each was mutation-checked: reversing the coalesce, dropping the queue-line read, and removing the backfill branch each fail the corresponding test.

Note on the second commit

Ownership of this PR transferred, and the mechanism was reworked. The first commit bound working_dir to the SessionWorker at creation time, which lost it on every restart (crash recovery and dead-letter replay create workers through call sites that had no access to it) and could never backfill, since ensure_session_node's node-exists branch returned early. The second commit moves the read to the queue line and adds the DB-level coalesce, and reverts what that made unnecessary: the get_or_create signature change, the HookStateService constructor param, and the per-Event-node stamp.

docs/architecture/03-graph-model.dot documents the new Session property; the PNG is regenerated.