Address deferred items from issue #48

[anilmurty]

Closes #48.

Three fixes, one per sub-item in the issue.

1. FastAPI lifespan migration (Greptile P2 from #47)

`tokenjam/cli/cmd_serve.py` had two `@app.on_event("startup" | "shutdown")` handlers — the legacy FastAPI API that's deprecated in newer releases. Converted to a single `@asynccontextmanager`-decorated lifespan and threaded it through `create_app(..., lifespan=...)`. `create_app()` grows an optional `lifespan` parameter (default `None`) so existing test callers keep working without changes.

Side benefit: `scheduler.start()` now runs inside the lifespan too — same reasoning as the existing comment about `server.state`: a failed-to-bind `tj serve` no longer spins up an orphan retention scheduler before the port-bind failure.

2. Doctor's drift-baseline check no longer warns on every new agent

`DriftConfig.enabled` defaults to `True`, so every brand-new agent (0–9 completed sessions) was tripping a `warning` on every `tj doctor` run. Collection-in-progress is the expected state, so this was just noise.

Changes in `_check_drift_inactive`:

• Downgraded `warning` → `info` for in-progress agents.
• Reworded as `Collecting baseline: agent (N/M)` so the line is informative rather than alarming.
• Multiple in-progress agents are reported in one line instead of returning at the first match.
• Threshold-reached agents stay silent (`ok`); agents with drift explicitly disabled are ignored.

Also tightened the spans-stats check's "skipped" messages — they now say "running through the HTTP API fallback" instead of the misleading "non-DuckDB backend". The backend IS DuckDB; the CLI session just doesn't have direct access while `tj serve` holds the write lock.

3. `cost_budget_daily` now fires in the budget breach demo

Diagnosed as a pricing-reference mismatch, not an alert-engine bug. The demo used model id `claude-sonnet-4-20250514` which isn't in `pricing/models.toml` (the table only carries the short-form ids `claude-{opus,sonnet,haiku}-4-X`). The cost engine fell back to default rates (`$0.50`/`$2.00` per MTok), making the 10-call total $0.0315 — below the $0.05 daily budget, so daily never breached. The session alert still fired because session budget is $0.02.

Switched the demo to `claude-sonnet-4-6` (which is in the table at $3/$15 per MTok), bringing total spend to $0.2205. Both alerts now fire as the demo's comments describe:

```
17:38:11 ⚠ CRITICAL cost_budget_session budget-demo Session cost $0.2205 exceeds budget $0.0200
17:38:11 ⚠ CRITICAL cost_budget_daily budget-demo Daily cost $0.2205 exceeds budget $0.0500
```

Not in this PR (worth tracking separately)

Adding dated-form Anthropic model ids (e.g. `claude-sonnet-4-20250514`) to `pricing/models.toml` is a broader concern — real users running real LLM calls will hit the same missing-pricing fallback and see surprisingly low costs. Worth filing as its own issue: either a fixture of common dated aliases, or normalization in `get_rates()` to strip the date suffix.

Test plan

• `pytest tests/unit/ tests/synthetic/ tests/agents/ tests/integration/` → 416 passed
• `ruff check tokenjam/` → clean
• `mypy tokenjam/` → "Success: no issues found in 78 source files"
• Live smoke test: `tj serve` starts cleanly under the new lifespan, writes `server.state` after port-bind, shuts down cleanly under `tj stop`
• `tj doctor` against a real DB: drift-in-progress reported as `i Drift detection: Collecting baseline: claude-code-tokenjam (0/10)` (info, not warning)
• Live re-run of the budget breach demo: both `cost_budget_session` and `cost_budget_daily` alerts fire

🤖 Generated with Claude Code