feat(telemetry): activation + reliability signals (schema v8)

[nicolotognoni]

Summary

• Closes the activation funnel and makes call failures diagnosable in the anonymous usage telemetry — the dataset previously showed that a call happened but not why an install never reached one, nor where latency/errors came from.
• Adds one event and several dimensions; all additive, opt-out, fire-and-forget (never affects a live call), no PII, Python ↔ TypeScript parity. Telemetry schema v7 → v8.

Implementation

• config_incomplete event — emitted once when required runtime config is missing, immediately before the existing startup raise (control flow unchanged). Dimension missing ∈ {carrier_credentials, llm_key, engine_config, other} — never the key value or any PII. This is the activation-blocker signal (installs that run but can't proceed).
• Per-stage latency on call_completed: stt_latency_ms, llm_ttft_ms, tts_first_byte_ms, eou_latency_ms (whole ms, read off the existing CallMetrics breakdown; a dimension is omitted when its stage didn't run, e.g. Realtime).
• Error taxonomy on call_completed: error_layer ∈ {stt,llm,tts,carrier,config,internal,none,other} + disconnect_reason ∈ {hangup_local,hangup_remote,error,timeout,no_answer,busy,completed,other}, derived from the already-available outcome/error_code (no new call-path plumbing).
• time_to_first_call_bucket on call_started/call_completed (install-age at call time).
• Client sampling: PATTER_TELEMETRY_SAMPLE (default 1.0 = no sampling), deterministic per run via the run id hash, never samples first_run/config_incomplete/error events; the chosen sample_rate is stamped so analytics can extrapolate (weight = 1/sample_rate).
• Files: libraries/{python/getpatter,typescript/src}/telemetry/{events,client,env,install_id,call_metrics}.* + client.*. New dedicated test files per feature (the shared telemetry test files are only touched for the SCHEMA_VERSION bump).
• The telemetry ingestion endpoint has been updated to accept schema v8, so the new event/fields are not dropped server-side.
• No version bump in this PR — entries sit under ## Unreleased; the bump + Unreleased → X.Y.Z rename happen at release time per the release workflow.

Breaking change?

No. Every addition is optional with a safe default (PATTER_TELEMETRY_SAMPLE defaults to 1.0), no existing event/dimension changed or removed, and the schema bump is purely additive. Existing callers are unaffected.

Test plan

• Python: pytest tests/ — telemetry 116 new/updated pass; full suite 2712 pass / 0 fail
• TypeScript: npm test (telemetry 87 pass; full 2142 pass / 0 fail) + npm run lint (clean) + npm run build (ok)
• Python ↔ TS parity verified (events, dimension value-sets, SCHEMA_VERSION=8, defaults all identical)
• E2E smoke — n/a (telemetry-only; no pipeline/handler behaviour change)

Docs updates

• CHANGELOG.md updated (## Unreleased → ### Added for the new signals, ### Changed for the schema bump).
• Follow-up (not in this PR): docs/telemetry.mdx to document PATTER_TELEMETRY_SAMPLE, the config_incomplete event, and the new call_completed dimensions; feature-inventory rows.