Update documentation with recent test stability improvements

Add details about a floating-point precision fix for boluswizardpreview.test.js and note an infrastructure issue with test process hangs during server teardown in replit.md.

Replit-Commit-Author: Agent
Replit-Commit-Session-Id: cf276ff5-4001-4b44-93fe-7c9786ee4327
Replit-Commit-Checkpoint-Type: intermediate_checkpoint
Replit-Commit-Event-Id: 0e1264f9-9ae7-4b11-a0e7-658fe3d1f1ac
Replit-Helium-Checkpoint-Created: true

docs/test-specs/flaky-tests.md

@@ -45,6 +45,24 @@ Some tests are slow due to server boot overhead (2-3s per test):
 
 ## Recently Fixed Tests
 
+### boluswizardpreview.test.js - Floating-Point Precision Fix (Fixed January 19, 2026)
+
+**Problem:** Test `set a pill to the BWP with infos` would intermittently fail, expecting `'0.50U'` but receiving `'0.51U'`.
+
+**Root Cause:**
+- The `roundInsulinForDisplayFormat()` function in `lib/sandbox.js` used `Math.floor(insulin / 0.01) * 0.01`
+- Floating-point precision errors caused values like `0.50499999...` to sometimes be represented as `0.5050000001...`
+- The floor operation at this boundary could produce either `0.50` or `0.51` non-deterministically
+
+**Fix Applied:**
+1. Added epsilon (`1e-9`) before floor operation: `Math.floor(insulin * 100 + 1e-9) / 100`
+2. Applied same fix to medtronic rounding style for consistency
+3. The epsilon is small enough not to affect normal values but stabilizes boundary cases
+
+**Verification:** Test passes 100% across 5 consecutive runs.
+
+---
+
 ### api.shape-handling.test.js (Fixed January 19, 2026)
 
 **Problem:** Test file was slow and occasionally timed out during stress testing due to excessive server boot overhead.

replit.md

@@ -143,11 +143,14 @@ npm run test:slow            # Tests with slow threshold logging
 **Summary:** ✅ **TESTS STABLE** - Stress testing (3-5 iterations per test file) shows 100% pass rate across all test files including api.shape-handling (now optimized).
 
 **Recent Fixes:**
+- `boluswizardpreview.test.js` - Fixed floating-point precision issue in `roundInsulinForDisplayFormat()` by adding epsilon (1e-9) before floor operation; test "set a pill to the BWP with infos" now stable
 - `api.shape-handling.test.js` - Fixed by optimizing server boot (beforeEach→before); now runs in ~6s with 172ms/test avg
 - `api.deduplication.test.js` - Fixed timeout issues by increasing timeout to 30s and optimizing cleanup
-- `boluswizardpreview.test.js` - All 10 tests now pass
 - `api3.renderer.test.js` - XML/CSV content type tests now pass
 
+**Known Infrastructure Issue:**
+- Full test suite runs may hang during server teardown (tick cycles continue after tests complete). Individual test files pass when run in isolation. This is a test cleanup issue, not test flakiness.
+
 **Flaky Test Status:** See `docs/test-specs/flaky-tests.md` for current status summary and improvement roadmap.
 
 #### Slow Tests