fix: read the renamed Otari-Request-ID and Otari-Attempt-ID response headers

[peteski22]

Why

Otari renamed its X- prefixed response headers to follow RFC 6648 (BCP 178), with no deprecation window, and that shipped in gateway v0.12.0 via mozilla-ai/otari#1667. This SDK read two of the old names, so both values come back empty against a current gateway.

The break is silent, which is what makes it worth fixing ahead of a report. _header_get returns None for a header that is not present, so with_response_metadata(...).request_id would simply be None on every response, and a failed call would lose its attempt identifier, with nothing raised for a caller to follow.

Codegen cannot carry this. The codegen workflow regenerates src/otari/_client/ only, and these header names appear in no generated artifact at all, which is tracked upstream as mozilla-ai/otari#1656.

What changed

Read Otari-Request-ID in place of X-Otari-Request-ID at four call sites, covering the sync and async clients, streaming and non-streaming. Read Otari-Attempt-ID in place of X-Correlation-ID in the error mapper, and carry that value as attempt_id rather than correlation_id, including in the error message a caller reads. That rename is more than the prefix: the header names the single provider attempt that served the request, while Otari-Request-ID names the whole resolve call containing it.

Both names are now module constants in _base.py, beside the existing GATEWAY_HEADER_NAME, rather than literals repeated at each call site. The README example names the new header.

No public surface changes. with_response_metadata and the streaming wrapper keep their current API, and correlation_id was never an exposed attribute here, only a local and a fragment of the error message text.

Notes

• Tests: UV_FROZEN=true uv run pytest: 139 passed, 7 integration tests skipped (no gateway). Each rename was driven test-first, and the red runs showed the exact silent failure: assert None == 'req-stream-123' across nine cases for the request ID, and a message of '[gateway] no funds' with the identifier absent for the attempt ID.
• Checks: UV_FROZEN=true uv run ruff check . passed; UV_FROZEN=true uv run mypy src/ passed (9 files). Both acceptance greps from the issue are clean: git grep -i "x-otari" and git grep -i "x-correlation-id" find nothing.
• Test literals: the unit tests keep spelled-out header names rather than importing the new constants, so a typo in a constant still fails a test. One test on each client keeps a lowercase spelling, holding the case-insensitivity coverage that _header_get provides.
• Release ordering: the issue requires this ship with or after the gateway release carrying Rename the released X-Otari-* headers to Otari-*, and write the header naming rule down otari#1485. That is v0.12.0, released 2026-09-25, so the ordering is satisfied.
• Existing lockfile mismatch: main's uv.lock records the root package as 0.2.0 while pyproject.toml is 0.4.0. Runs used UV_FROZEN=true to avoid an unrelated lockfile edit, and test: cover container serialization in messages #94 already carries that sync.
• Not covered here: Otari-Guardrails, Otari-Container-Id and Otari-Container-Expires-At were renamed in the same gateway change but are not read by this SDK.

Fixes #96