feat(model-gateway): X-RLLM-* headers + inbound bearer auth

[jeffreysijuntan]

Summary

Stacked on #551. Header-stamped session metadata becomes the canonical identity carrier; body fields and the legacy /sessions/{sid}/v1/... URL form are kept as field-by-field fallbacks. rLLM headers are stripped before forwarding upstream so they never leak to providers.

• New metadata.py: RllmMetadata dataclass + extract_metadata() with precedence header > body > path, merged field-by-field (so a request can carry session_id in the URL and run_id in a header and both contribute).
• Middleware attaches RllmMetadata to ASGI scope and gains an optional inbound-auth gate: when a bearer token is configured, every request must present Authorization: Bearer <token> (HMAC-compared) or get 401. Used by remote-sandbox eval where the gateway is exposed via a public tunnel.
• TraceRecord gains the rLLM identity fields (run_id, harness, step_id, parent_span_id, span_type); data_process.build_trace_record populates them from RllmMetadata when present.
• models.py carries the future-facing config fields (UpstreamRoute, routes, run_id, run_metadata, inbound_auth_token) so the upstream-proxy PR can wire them up without further model churn.

Stack

• PR-1A — schema v2
• [PR-1B] this — header metadata + inbound auth
• PR-1C — upstream-proxy mode (depends on this)

Test plan

• python -m pytest tests/unit/ — 242 passed
• 15 new tests in test_metadata.py (header/body/path precedence, field-level merge, scope-shape decoding)
• 7 new tests in test_inbound_auth.py (token absent → open, presented → 200, missing/wrong → 401, HMAC compare)

🤖 Generated with Claude Code