Compass / Engineering review
7 September 2026 · Overnight work

Clearer imports.
Safer test runs.

A complete change log for the work completed while you slept, including the defects caught during review and the evidence behind each fix.

Verified locally · ready for review
13documented improvements
2,609examples in the final suite
0failures in final verification
6synthetic browser screenshots

What changed

Each item states the resulting behavior. Open its evidence for the previous behavior and the checks that cover it.

01 / Imports

Import action totals

Live summary distinguishes selections, uploaded values, kept rows, exclusions, invalid rows and conflicts.

Before and evidence

Before Checked Keep/Skip rows were described as imports.

Evidence Mixed-action browser tests assert saved records.

02 / Imports

Accurate conflict detection

Compare the typed values the importer saves, including sell currencies and exact free text.

Before and evidence

Before Identical matches inflated conflicts; sell-currency-only and text-case changes could be hidden.

Evidence Request regressions.

03 / Imports

Complete comparisons and errors

Show full uploaded/existing values and visible invalid-row explanations.

Before and evidence

Before Long values were truncated; errors depended on tooltips.

Evidence Request and browser checks; synthetic screenshots.

04 / Imports

Accessible row controls

Named controls, a live summary and a labeled keyboard-scrollable comparison region.

Before and evidence

Before Row inputs lacked useful names.

Evidence Keyboard browser check.

05 / Imports

Usable long reviews

Sticky wrapping footer, contained table scrolling and a phone scroll hint.

Before and evidence

Before Confirmation scrolled out of view; phone containment needed repair.

Evidence Desktop 1400 and phone 390 checks in normal and replace-all modes.

06 / Imports

Explicit replace-all scope

Show records removed, selected replacements and omitted-record consequences; restore Keep/Skip choices when toggled off.

Before and evidence

Before Warning omitted existing counts and did not preserve earlier row choices.

Evidence Browser persistence and mode-toggle tests.

07 / Imports

Document currency restrictions

Request, retain and validate the field through extraction and confirmation.

Before and evidence

Before Extraction discarded an explicit allowed sell-currency list; confirmation could clear a saved restriction.

Evidence 5 new examples, 4 failed before the fix; extraction group 80/0 afterward.

08 / Development

Stable test class identity

Disable test reloading and bypass Spring together.

Before and evidence

Before Watched-file changes could reload Rails classes during a suite.

Evidence Real automatic-reloader subprocess fails under old configuration and passes under new.

09 / Development

Guarded isolated test command

Explicit numeric database suffix, preboot/resolved-config guards, database and checkout locks through process exit, explicit safe preparation and normal RSpec discovery/status.

Before and evidence

Before Environment overrides or concurrent runs could target shared state.

Evidence 31 runner examples plus an automatic-reloader regression; two real repeated database preparations.

10 / Development

Accurate setup and documentation

Setup follows the actual Ruby/Rails asset stack; document real test, extraction and background-worker requirements.

Before and evidence

Before Setup required nonexistent Yarn dependencies and included unwanted seeding/cleanup/restart behavior; docs described the wrong stack.

Evidence 95 existing request/model/job/browser examples through the guarded raw RSpec path.

11 / URL transport

Pinned URL destination

Connect to validated addresses while retaining the original TLS and Host identity; disable environment proxy bypass.

Before and evidence

Before DNS validation and socket connection resolved separately.

Evidence Socket-level regressions and two real local TLS handshake tests, including wrong-host certificate rejection.

12 / URL transport

Special-use address checks

Normalize mapped addresses and reject blocked destinations before HTTP creation.

Before and evidence

Before Mapped private addresses and additional special-use ranges escaped checks.

Evidence Blocked/mixed-address regressions.

13 / URL transport

Safe address fallback

Try only captured validated addresses under the existing total deadline, including unsupported-address-family failures.

Before and evidence

Before Initial pinning could fail a valid dual-stack URL at its first unreachable address.

Evidence Connection failure, address-family and deadline regressions; 81 existing service examples plus two TLS examples pass in the integrated suite.

Showing all 13 improvements.

Verification

Counts below are executed examples. Zero-example exits and failed runs are recorded separately and never counted as passing verification.

CheckResultEvidence
Fixed baseline2,537 examples · 0 failuresSeed10459; database 201.
Reviewed full suite2,609 examples · 0 failuresSeed 10459 at reviewed head 471337daa.
Different test order2,609 examples · 0 failuresSeed 83719 at reviewed head 471337daa.
Final committed suite2,609 examples · 0 failuresSeed 83719 at final application head 93e67d028; focused follow-up also passed 88 examples.
Focused integration201 examples · 0 failuresImports, extraction and runner regressions after simplification.
Real database preparationTwo repeated runs passed27 examples each; selected database 203; postgres catalog unchanged.
Ruby style407 offenses; baseline 414Repository-wide lint remains red. All five new Ruby files have zero offenses.
Independent review8 of 8 lenses completedAstra ultra, plus a separate validator; one P2 test finding applied in 93e67d028.
Report browser checksDesktop and phone passed1400 px and 390 px: filters, details, six images and page containment.

What review caught

Independent Astra reviews and runtime checks found issues in the first implementations. These follow-ups are part of the final branch.

  • Existing test-database preparation could leave Rails connected to postgres. The runner now restores and verifies its intended database before loading a schema.
  • Test locks could release before older exit hooks finished. File descriptors now stay open until process teardown.
  • A pathless test command initially ran zero examples and returned success. The launcher now preserves RSpec’s command name; real CLI regression cases prove discovery and failure status.
  • The first DNS pinning implementation stopped at an unreachable address. Fallback now uses only prevalidated addresses, including address-family failures, under one total deadline.
  • Mobile review initially overflowed the page. The table now scrolls within the phone viewport and the confirmation controls remain usable.
  • Simplification removed an extra COUNT query and reused the parser’s field definition without sharing mutable field strings.
  • Adding document sell-currency restrictions exposed a stale VCR header expectation in the first full integration. The expectation was corrected without recording an external request.
  • The final review found a test that relied on a real 10 ms window. A controlled monotonic clock now expires only during the required first socket attempt; the test explicitly forbids a second address.

Browser evidence

Desktop and phone views use synthetic data. Select an image to inspect the original capture at full size.

Review the work locally

One isolated branch

feat/overnight-compass-20260907

Final application commit 93e67d028, based on 7d64d2ba1. The review receipt records its original scope and the applied follow-up.

Review the combined implementation in its own checkout. The report and its synthetic screenshots are the only published assets.

cd /home/stevan/dev/compass-overnight-20260907
env -u DATABASE_URL -u PRIMARY_DATABASE_URL RAILS_ENV=test RACK_ENV=test TEST_ENV_NUMBER=201 DEVISE_SECRET_KEY=synthetic-overnight-test-key bin/test --seed 83719

Scope and practical limits

  • Application changes are local only. Nothing pushed, merged into main or deployed as an application.
  • Authentication/email and lookup/evaluation were reserved for the other session. This is additional work on the merged baseline, not a new verification of every earlier finding.
  • No migration or dependency added. Live model extraction was not exercised; transport behavior was verified with controlled local fixtures.
  • Repository-wide lint is still red. The report records its exact count and the new-file result separately.
  • Worker worktrees and databases 201–204 remain available for inspection.
  • Three advisory test extensions are recorded in the independent review: first database creation, other upload-type comparisons, and Stimulus reconnection.