Dashboard report snapshots per scope

[SaulMoro]

Parent

#752 (item 1, remainder)

What to build

Dashboard report snapshots are kept per scope. The report sends deltas: for each session it compares the current totals with the last reported ones, kept in machine-level reported-*.json snapshots keyed by session. Once #785 is done, no scope reports another scope's sessions. One case still breaks: a session whose events belong to two scopes, for example after a cd into another project mid-session.

session X: events in scope A (3 prompts) and scope B (2 prompts)

shared snapshot                         per-scope snapshot
A reports 3 → snapshot[X] = 3           A: snapshot_A[X] = 3
B: delta = 2 − 3 → wrong                B: delta = 2 − 0 = 2

Move the snapshots next to each scope's data. The first time a scope needs its snapshot, it seeds it from the shared one, so nothing already reported is sent again. After that, the shared file is not read. The user scope's data home is the machine-level directory, so it keeps the existing file.

Acceptance criteria

• A session split across two scopes is reported correctly to both teams.
• The first pull after the upgrade sends no delta for sessions that were already reported.
• A scope's later pulls read and write only its own snapshot.
• Verified with the real CLI (Claude; git, gitlab, github providers); data-directory design and CHANGELOG updated.

Blocked by

• Each scope reports only the dashboard sessions recorded in it #785

[SaulMoro]

Design correction found while implementing: the user scope must not keep writing the shared reported-*.json. If it does, a project seeds from values the user scope reported after the upgrade, and the split case this issue fixes stays broken:

session X: user scope 3 prompts, project P 2
user-scope pull after upgrade → shared[X] = 3
P's first pull → seeds P[X] = 3 → delta 2 − 3 → 0     P never gets X

So every scope writes only its own snapshot. The user scope moves to ~/.teamai/dashboard/user-reported-*.json (same precedent as user-usage.jsonl in #758) and seeds from the shared file on first use like any project. The shared reported-*.json is never written again and serves only as the frozen seed.

[SaulMoro]

Design change, decided while fixing the CI review on #791 (which now carries this issue): a Stop event carries the transcript's cumulative totals, so a session cannot be split between scopes event by event without over-counting (user 3 prompts, then P 2 → P would get 5). A session is now reported once, whole, to the scope of its first keyed event. Per-scope snapshots stay, to keep sessionId collisions (Copilot's PID fallback) apart. The acceptance criterion "a session split across two scopes is reported correctly to both teams" becomes "reported once, to the scope it started in".

[SaulMoro]

Two deviations from this issue, accepted by a maintainer in #791:

1. "A session split across two scopes is reported correctly to both teams" → a session is reported once, whole, by the scope it started in. A Stop carries the transcript's cumulative totals, so B's Stop in the example reports 5, not 2. Splitting per event would make B count A's 3 again. The cost: a session that starts in one scope and continues in another counts only in the first.
2. "A scope's later pulls read and write only its own snapshot" → reports still read the shared snapshots, to tell main's copies of shared entries apart from a scope's own reports. Every scope also writes the machine-level ~/.teamai/dashboard/session-owners.jsonl (a session ID and a hashed data-home key per line, never a path), which keeps a session resumed after compaction with the scope that reported it.

The first pull after the upgrade still sends nothing already reported, and the per-scope snapshots are as specified.