fix(telemetry): deliver call_completed before process exit via drain() in disconnect()

[nicolotognoni]

Summary

• E2E-testing the schema-v6 telemetry with a real outbound phone call exposed a data-loss bug: call_started reached Axiom but call_completed — the one event carrying duration/cost/latency/outcome — never did, for every short-lived script ("place call, wait, exit": the main outbound use case). The connect→complete funnel read as a 100% failure rate.
• Root cause (Python): disconnect() used a fire-and-forget flush; the process exited right after, asyncio.run() closed the loop cancelling the flush task mid-POST, and the 0.25s atexit fallback was too short for a cold TLS handshake. (This confirms the "atexit 0.25s too short" P2 from the hardening audit — now field-proven, with a worse blast radius than assumed.)
• Fix: new TelemetryClient.drain() in both SDKs — flush buffered events and await delivery (bounded), keeping the client reusable (unlike aclose()/close(), a subsequent serve() still emits). disconnect() now awaits it. Python's atexit timeout bumped 0.25 → 1.0s as a second line of defense for scripts that never call disconnect().

Implementation

• libraries/python/getpatter/telemetry/client.py — drain(timeout) (awaits in-flight flush task + flushes remainder, swallow-all); _ATEXIT_TIMEOUT_S 0.25 → 1.0.
• libraries/python/getpatter/client.py — disconnect() awaits self._telemetry.drain() (was fire-and-forget flush_pending()).
• libraries/typescript/src/telemetry/client.ts — drain() mirror (TS was structurally safer — an in-flight fetch keeps the Node loop alive — but now mirrors Python for explicitness and parity).
• libraries/typescript/src/client.ts — disconnect() awaits this.telemetry.drain().
• Authentic tests in both SDKs (real local HTTP collector, no mocks): drain delivers the final event AND the client stays usable afterwards.
• Test rename: test_build_event_v5_new_events_and_dims → ..._v6_... (stale name after the schema bump, flagged by review).

Field evidence (real calls, Axiom dataset getpatter_usage)

• Run 1 (pre-fix, 13:21Z): call answered, 32s, CallResult complete locally — call_started in Axiom, call_completed missing (3+ min wait).
• Run 2 (post-fix, 13:34Z): call answered, 25s — call_completed landed <1s after disconnect with full metrics: outcome=completed, duration_seconds=25, cost_usd=0.06, turn_count_bucket=4_6.

Breaking change?

No. drain() is additive; disconnect() gains a bounded (≤2s, typically ~100ms) wait that only runs when telemetry is enabled and events are pending. Opt-out paths unchanged.

Test plan

• Python: pytest tests/ — 2656 passed, 8 skipped, 2 xfailed
• TypeScript: npm test (2102 passed) + npm run lint + npm run build — green
• /parity-check — drain() same name/semantics both SDKs, disconnect() parity
• E2E smoke — two REAL outbound calls (see field evidence above)

Docs updates

• N/A (internal telemetry delivery mechanics; docs/telemetry.mdx event semantics unchanged)

Known follow-ups (out of scope, noted during E2E)

• direction present on call_started but empty on call_completed (call-end payload doesn't propagate it).
• TS SCHEMA_VERSION not exported from the package root (internal-only).