perf(users/test): make bcrypt cost cheap under go test

[dvcdsys]

What

Password hashing uses bcrypt cost 12 (~250 ms/hash, by design). Almost every
server test seeds a user through a fixture, so that deliberate cost dominated the
whole -race suite. This resolves the work factor once at init and drops it to
bcrypt.MinCost under go test.

Why

Per-package go test -race ./... times were mostly pure bcrypt, serialized and
amplified by CI CPU contention:

package	before (CI, -race)	after (local, -race)
httpapi	462.3 s	15.8 s
users	83.0 s	<1 s
sessions	43.3 s	<1 s
apikeys	27.9 s	5.4 s
chunker	22.9 s	18.9 s (untouched)

Full suite after the fix: all green, 0 data races, ~23 s wall locally.
(chunker is tree-sitter/CGO — inherently slow under -race, left as-is.)

How

internal/users/users.go — BcryptCost goes from a const to a value resolved
once at package init, highest precedence first:

1. CIX_BCRYPT_COST — explicit override, clamped to [bcrypt.MinCost, MaxCost].
2. testing.Testing() — under go test, use bcrypt.MinCost (~256× cheaper).
   The hash is never what the tests assert; round-trip behaviour is identical.
3. 12 — production default (unchanged outside tests).

Safety: testing.Testing() is false outside test binaries, and since Go 1.13
importing testing no longer registers test flags, so cix-server's
flag.Parse() is unaffected — verified cix-server -v still works. We're on
Go 1.25.

Also derived the user-not-found anti-enumeration dummy hash from the active cost
(was a hard-coded $2a$12$… literal) so the timing mitigation tracks the cost
and the not-found path is cheap under test too.

Type of change

• Refactor (test performance; no production behaviour change)

Checklist

• go test -race ./... passes — all green, 0 data races
• go vet ./internal/users/ passes
• cix-server builds and -v runs (testing import doesn't break flag parsing)
• No secrets or API keys committed

🤖 Generated with Claude Code