fix: Claude Code session status + lifecycle (1/5 from #306)

[anshss]

Carved from #306 as PR 1 of 5 — a focused split of the session-status dashboard branch. This one delivers the root bug fix + the honest session lifecycle model it enables.

Root fix

Claude Code / Codex export OTLP logs, not traces — so each user_prompt turn produces a zero-duration invoke_agent span (end_time == start_time). ingest mistook that for a session completion, so every live session force-completed on its first prompt (dashboard showed active work as "completed" with 0 duration, and it tripped drift/alert false-positives). Completion is now gated on real duration (end_time > start_time); a non-end span re-activates a mistakenly-completed session.

Lifecycle model (built on the fix)

• Tiers active / idle / stale / closed via SessionRecord.status_at() (5 min stale, 4h idle default), with a transcript-mtime rescue for live-but-span-quiet sessions.
• Explicit close — POST /api/v1/sessions/close + tj session-end (preserves ended_at / "Last seen"; only stamps status).
• Project grouping — OTel service.namespace + a server-side agent→project fallback (groups already-running sessions without a restart).
• Display labels — service.instance.id + manual overrides (right-click rename) + tj otel-resource-attrs.
• Split-zones Status — coding sessions (cards + archived) vs SDK services (per-minute cost/err sparklines); 0-signal "zombie" sessions filtered from the archive.

Migrations

8 service_namespace · 9 service_instance_id · 10 ended_at repair · 11 session_labels (renumbered contiguous from main's current max of 7).

Notes for the series

• Decoupled from later PRs: the capture_session_method snapshot on close, the session_token_cost_rollup (Polish web UI for demo-readiness #18) used by the zombie filter, and the core.transcript mtime-rescue are all deferred — PR1's close endpoint just closes; the zombie filter uses stored token/cost. Polish web UI for demo-readiness #18 lands in the cost PR; method-capture in the Approach/Map/Timeline PR.
• uv.lock (introduced on the branch, untracked on main) is intentionally not included here.

Tests

Full suite green: 1400 passed, 1 skipped (pytest tests/unit tests/synthetic tests/agents tests/integration). ruff check clean; index.html module validated with node --check.

🤖 Generated with Claude Code

[anilmurty]

Reviewed the full lifecycle fix — no blockers from me. Root diagnosis is right: the logs path emits zero-duration invoke_agent markers every turn, and gating completion on end_time > start_time plus re-activating on later activity is the correct fix. Ran the touched suites locally, green. Rules clean — SQL is $N-bound, the new /sessions/close write endpoint is in PROTECTED_PATHS, migrations 8–11 idempotent. Three tiny non-blockers: status_with_transcript_mtime ships unused (fine as scaffolding), _serialise writes a [sessions] block on every config save, and the shell wrapper double-fires session-end on interrupt (idempotent, harmless). This is the base of the stack, so it merges first.

[anilmurty]

Approving per the review above — root lifecycle fix is sound, rules clean, suite green. Merging as the base of the stack (#371 follows).