fix(credential-health): stop the sweep defaulting to 5 minutes

[adivekar-utexas]

What

#12138 raised the credential health sweep default to 60 minutes, but left the 5-minute value behind in four places. The sweep can therefore still run at 5 minutes — about 12x the intended cadence — issuing a real credential probe against every active connection.

The leftovers

Location	Was	Now
getConnIntervalMs(conn, globalIntervalMs = 300_000)	5 min default param	shares DEFAULT_SWEEP_INTERVAL_MS
open-sse/config/constants.ts CREDENTIAL_HEALTH_CHECK_INTERVAL	300_000	3_600_000
docs/reference/ENVIRONMENT.md	default 300000	3600000
src/lib/copilot/systemPrompt.ts config table	300000	3600000
scheduler.ts header comment	"default 5 min"	"default 60 min"

The first is the only functional one, and it is latent rather than live: both current call sites pass globalIntervalMs explicitly. It is exactly the trap that resurfaces the moment someone adds a third call site.

Why a shared constant rather than a corrected number

DEFAULT_SWEEP_INTERVAL_MS now lives in the probePolicy leaf, which scheduler.ts already imports (so no cycle), and both the global resolver and the per-connection fallback read it. The two cannot drift again.

The change inside resolveInconclusiveProbeRecheckDelayMs is behaviour-neutral: its only caller passes getSweepInterval(), which is always finite and positive, so that fallback branch is unreachable today.

Tests

Three cases added to tests/unit/credential-health-sweep-interval.test.ts:

• the per-connection fallback is 60 minutes, not 5
• healthCheckInterval: 0 opts the connection out entirely
• a per-connection interval still overrides the global cadence

18/18 pass across the credential-health suites. check:env-doc-sync, check:file-size, typecheck:core, check:open-sse-typecheck and ESLint are all clean.

For operators landing here from probe traffic

If you are seeing frequent tiny completions against a connection, the sweep is the cause. Three levers, in precedence order:

1. Per-connection — Edit Connection → "Health Check (min)" → 0 (never test this connection)
2. Global — Settings → Resilience → credential health check → 0 disables the sweep
3. Env — CREDENTIAL_HEALTH_CHECK_INTERVAL (ms, minimum 10000)

A stored intervalMinutes or healthCheckInterval overrides the 60-minute default, so if you are on 5 minutes today that is almost certainly an explicit value left behind from testing #12043 / #12138 rather than the default.