eidolons v1.41.0

Telemetry P2 sprint 2 — budget gate + store export

Builds on the v1.40.0 telemetry subsystem.

telemetry budget — cross-run budget gate

eidolons telemetry budget --limit N [--by repo|branch|model|eidolon|tier|day] [--since DATE] [--usd] [--json]. Aggregates spend over the telemetry store, grouped by --by (default eidolon), source-split (audited/estimated evaluated independently — never summed). Exits 3 when any group exceeds the limit (mirrors trace cost --budget), 0 when within. --usd evaluates a dollar threshold via roster/pricing.yaml, with an honest "can't evaluate $ budget for unpriced model" note rather than a silent pass. --json for alerting pipelines.

Real-data example: budget --usd --by eidolon --limit 5 over a delegation-heavy session → exit 3, main: $247.15 > $5 BREACH, with the per-Eidolon dollar breakdown (atlas $0.07 · forge $0.25 · spectra $0.34 · vivi $0.40).

telemetry export — store export

eidolons telemetry export [otel|json|csv] [--since DATE] [--project <slug>], stdout-only (the nexus bundles no collector — pipe to any backend). otel emits OpenTelemetry GenAI-convention spans reusing the SAME gen_ai.*/eidolons.* attribute keys as trace otel (single source of truth); json emits the deduped store array; csv flattens the core columns (incl. usd when priced) with RFC-4180 quoting.

Notes

Additive, opt-in, bash 3.2 + jq + yq only. 41 new tests; full suite green on ubuntu + macOS. Roadmap update: the M2 reconciliation metric (P2.3) was dropped after a feasibility scout — no machine-measured self-report data exists to reconcile against (the report's honest "N/A" is the correct end-state). Remaining P2 in .spectra/telemetry-tooling-2026-06/p2-roadmap.md.