Release 0.3.0

[anassg-lago]

Cuts 0.3.0 for PyPI. Version bump + CHANGELOG.md [Unreleased] → [0.3.0] - 2026-09-02; no source changes.

Merging this does not publish. The release is triggered by pushing the v0.3.0 tag afterwards, which runs publish.yml and then pauses on the pypi environment for maintainer approval before anything reaches PyPI.

What 0.3.0 contains

• Ramp Router support — a wrapped OpenAI client pointed at Router previously billed nothing at all; detected from the base_url host, billed on the model that actually served the call.
• Snowflake Cortex connector — live wrapper + REST/METRICS backfill, with x-snowflake-request-id keying so a backfill of an already-billed window is rejected by Lago instead of double-billing.
• Databricks connector — AI Gateway usage + spend backfill.
• The total_tokens guard now reads each provider's own convention from the new token_semantics.py, after Cortex was measured answering on an OpenAI wire with Anthropic's additive convention (2.0x on the call, ~800x on the output line).
• Queue and pricing fixes: connection reuse, honest flush(), 4xx reclassification, adapter errors reaching on_error, pricing TTL + per-source backoff.

Full detail in CHANGELOG.md.

Gates

Run locally on this commit: ruff + mypy strict + 835 tests green; CI re-runs the matrix on the PR.