feat(security): prevent user enumeration via timing attacks

[maxdinech]

What kind of change does this PR introduce?

Feature - Security enhancement

What is the current behavior?

Password authentication endpoints leak information about user existence through timing differences:

• Valid users (wrong password): ~120-150ms (includes DB eager loading + bcrypt)
• Invalid users: ~10-20ms (fast DB lookup, no bcrypt)

This timing difference allows attackers to enumerate valid email addresses/phone numbers by measuring response times, even without rate limiting bypass.

We discovered this vulnerability during a penetration test of our self-hosted Supabase instance. Testing showed a consistent 100ms+ difference that reliably identified valid accounts.

What is the new behavior?

Adds a new configuration option GOTRUE_SECURITY_TIMING_OBFUSCATION_DELAY that adds timing obfuscation to invalid login attempts.

Changes:

• Added GOTRUE_SECURITY_TIMING_OBFUSCATION_DELAY configuration option
• Invalid user lookups now perform dummy bcrypt comparison (same cost as real passwords)
• Configurable additional delay compensates for database eager loading overhead
• Applied to both "user not found" and "user has no password" cases

Implementation details:

• Uses pre-computed bcrypt hash (cost 10) for consistent timing
• Bcrypt operation auto-scales with CPU speed (~50-200ms)
• Additional delay compensates for .Eager() loading of Factors/Identities (~30-50ms)
• Recommended value: 40ms (empirically tested to overlap timing ranges)
• Disabled by default (opt-in security enhancement)
• No database schema changes required

Testing performed

Timing analysis on staging environment (20 samples each):

• Before fix: 113ms average difference (VULNERABLE)
• After fix (40ms delay): 7ms average difference with overlapping ranges (PROTECTED)

Valid user and invalid user response times now overlap significantly, preventing reliable enumeration.

[maxdinech]

Note : on more powerful instances, a value of just 10ms for GOTRUE_SECURITY_TIMING_OBFUSCATION_DELAY can be sufficient. The main introduction of that PR is to force some bcrypt call even on the case where the user is not found.