test(pool): concurrent-client HTTP load test for the DB connection pool

[tcconnally]

Closes #223.

Follow-up to the connection pool (#210, shipped in #221). The existing unit tests (pooled_database_shared_across_threads, concurrent_reader_writer_no_locks) exercise Database directly — not the HTTP/SSE transport — and the pool knobs were hard-coded, so the durability/throughput characteristics under sustained concurrent client load were unverified.

Changes

1. Make the pool tunable via env (src/db.rs)

• MIMIR_POOL_MAX_SIZE (default 16)
• MIMIR_BUSY_TIMEOUT_MS (default 5000)

Defaults preserve prior behavior. This lets operators size the pool to their workload and lets the load test sweep the knobs the issue asks about.

2. Add an #[ignore]d load test (src/transport.rs)
Drives the real HTTP transport — the same init_transport_state + build_transport_router + axum::serve path main.rs uses — with N concurrent ureq clients interleaving:

• writes: mimir_remember with unique high-entropy bodies (so each is a real create, not a near-duplicate dedup — otherwise persisted == issued wouldn't test durability)
• reads: mimir_recall + mimir_context

It asserts the four properties #223 calls out:

• ✅ no database is locked / SQLITE_BUSY after the busy_timeout
• ✅ no lost writes — rows persisted (counted via an independent connection) == writes issued
• ✅ no deadlock — all client threads join
• ✅ reports p50 / p99 / max latency + throughput so the operator can judge tail behavior

Why #[ignore]

It's a load/soak test, not a CI correctness gate — the contention characteristics "can't be proven by CI" (per the issue). Run it explicitly and sweep:

cargo test --release pool_load_test_http_transport -- --ignored --nocapture

# sweep the pool knobs
MIMIR_POOL_MAX_SIZE=4 MIMIR_BUSY_TIMEOUT_MS=5000 MIMIR_LOADTEST_CLIENTS=32 \
  cargo test --release pool_load_test_http_transport -- --ignored --nocapture

Tunables (env): MIMIR_LOADTEST_CLIENTS (16), MIMIR_LOADTEST_WRITES / MIMIR_LOADTEST_READS per client (25 / 75).

Verification (x86_64-pc-windows-msvc, debug)

Default 16 clients × 16 pool: 2800 requests, 400/400 writes persisted, 0 lock errors, 0 other errors, p50≈1ms / p99≈27ms. The default cargo test run shows it ignored (does not gate CI); existing pool/transport tests still pass.

🤖 Generated with Claude Code