[DarknessKiller]

Three targeted performance/stability changes, each in its own commit, all verified by build/vet/test plus local A/B evidence.

Changes

1. HTTP connection reuse across probes (internal/service/uptime)

• Before: every HTTP probe built a fresh resty client - new connection pool, full TCP(+TLS) handshake, tore down after each request.
• After: one long-lived client per (resolver, TLS-verify-mode) on the Service; probes dial over pooled, keep-alive connections; per-host TLS NoVerify and per-DNS dialer behavior unchanged.

2. SQLite WAL journaling (bootstrap)

• Appends journal_mode(WAL) + synchronous(NORMAL) to the DSN; the libtnb driver already defaults busy_timeout(5000).
• Readers no longer block during probe writes; cheaper fsync per probe.
• Note: keep the SQLite file on local or networked-local persistent storage - WAL breaks on network filesystems.

3. Vendor chunk split (frontend)

• react/react-dom/scheduler/axios build into a stable vendor-<hash>.js chunk.
• On app deploys, browsers reuse the cached vendor chunk (26 kB index + 237 kB vendor vs one 288 kB blob).
• HostDetail/NotificationManager were already lazy imports; g2plot stays out of the initial payload.

Local A/B evidence (same load shape, single variable = shared pool)

Side	Probes	TCP connections at target	conn/probe
before	110	205	~1.86
after	100	5	~0.05

Verification

• go test ./... incl. new TestService_HTTPClientReuse and TestSQLiteDSNPragmas
• go vet ./... clean, go build ./... clean
• vite build artifact sizes recorded above