feat(model-gateway): trace store schema v2

[jeffreysijuntan]

Summary

First of three stacked PRs splitting the gateway upgrade. Standalone — no consumer changes; landed alone, the new schema and methods are dead code that the next two PRs activate.

• Denormalises hot filter columns (run_id, model, harness, latency_ms, has_error, step_id) onto the traces row so the cross-run trace feed (rllm-console PR) can filter without JSON-scanning every row.
• Adds a runs table for per-gateway-run lifecycle.
• Extends TraceStore protocol: query_traces, register_run, end_run, list_runs, count_traces, facets.
• Schema-version mismatch on an existing db drops & recreates every table — gateway data is dev-only and a real migration isn't worth it.
• _get_conn serialises first-call init under an asyncio.Lock so a burst of concurrent store_trace callers can't race the schema build (the previous lock-free path could leave racers staring at a half-dropped db when migration triggered).
• Memory store mirrors the new shape so unit tests cover both backends identically.

Existing store_trace(trace_id, session_id, data) callers keep working — run_id defaults to "", all new methods are additive.

Stack

• [PR-1A] this — schema v2 + lock fix
• PR-1B — X-RLLM-* headers + inbound bearer auth (depends on this)
• PR-1C — upstream-proxy mode + run lifecycle (depends on B)

Test plan

• python -m pytest tests/unit/ — 220 passed
• python -m pytest tests/unit/test_store.py — 74 passed (new tests for schema v2, denormalised columns, junction back-compat, concurrent writers)
• One known flaky case: test_two_stores_share_one_db can fail ~5% under suite-wide ordering when two SqliteTraceStore instances race the WAL pragma. Passes reliably in isolation. Tracking as a follow-up — fix is a small retry on journal_mode=WAL or a process-level init lock.

🤖 Generated with Claude Code

[luyuzhe111]

@jeffreysijuntan look good overall. what's the PR that tackles the flaky test tho?

[jeffreysijuntan]

@jeffreysijuntan look good overall. what's the PR that tackles the flaky test tho?

It is in a new uncommit PR, I will push it to here instead

[listar2000]

What's the current blocker for this PR to get merged (the CI failure?)

[jeffreysijuntan]

What's the current blocker for this PR to get merged (the CI failure?)

Before changing the gateway, I am working on some PR to unify the eval path with the AgentFlowEngine, so eval will run directly through that in the future.