Split mixin loading across system and integration tiers

[mkarlesky]

Summary

PR 4 of 5 in the system-test redesign (Stage 2 plan, following the "System Test Duplication Assessment" report's Finding 6) -- the second-largest lever, an estimated ~5.3% of measured system-test time.

mixin_ordering_spec.rb tested precedence/dedup/ordering/sigil/:tools-append merge mechanics via a full ceedling CLI round-trip per scenario, even though that logic is pure config-hash transformation with no CLI or OS dependency -- and already has 1,970 lines of dedicated unit coverage across composinator_spec.rb, mixinator_spec.rb, mixin_resolvinator_spec.rb, recursive_merger_spec.rb, mixin_standardizer_spec.rb, and merginator_spec.rb. This is a gap left over from when system tests were the only tier available, not a design choice.

Three-way split, not a flat migration

• spec/system/mixin_ordering_spec.rb: 36 → 7 examples. Kept only what's genuinely CLI-surface with no unit-tier analog: does --mixin/CEEDLING_MIXIN_*/a project's own :mixins config section actually get read from the real command line/environment/file, and does the pipeline exit non-zero on bad input.
• New spec/integration/mixin_loading_spec.rb: 17 examples. A real Composinator/Projectinator/MixinResolvinator/Mixinator/MixinStandardizer/Merginator/RecursiveMerger graph, real files on a real temp directory -- no CLI subprocess, no gem deployment.
  • 10 net-new: project-directory-as-load-path (3), built-in vendor/unity/test/targets/ resolution (6), :extension yaml override (1) -- real filesystem resolution no unit spec exercises (they all stub file_wrapper/load_paths away).
  • 7 representative wiring smoke, consolidated from 19 system-tier examples whose logic is already proven exhaustively by the six unit specs above: one each for dedup, merge order, duplicate --mixin ordering, overall precedence, three-way inline-YAML positional order, the @ sigil, and the :tools arguments append exception -- proving the real object graph produces the same result end to end, not re-verifying logic those unit specs already cover scenario by scenario.

config[:history][:config] (Mixinator#mixin's own structured merge record) is used throughout the new integration spec in place of the old spec's raw log-text scanning -- a strictly more precise, directly-inspectable signal for dedup/ordering than grepping console output.

Verification

• New integration spec: 17/17.
• Trimmed system spec: 7/7.
• Full rake specs:integration: 102/102 (85 existing + 17 new).
• The six untouched mixin unit specs: 117/117 (confirms this PR touches no unit-tier files).
• All green, host (macOS) and the madsciencelab-plugins Docker image (Linux).
• Full-suite validation (all 16 CI matrix jobs) is CI's job per the Stage 2 plan's process -- not duplicated locally.

This is PR 4 of 5; the plan calls for a go-ahead before starting PR 5 (example-project audit + new files:/paths: spec).

🤖 Generated with Claude Code

[mkarlesky]

Coverage investigation: why combined coverage dropped slightly

Downloaded both PRs' coverage-report CI artifacts and diffed SimpleCov's per-line hit data directly (PR #1288's run vs. this PR's run) rather than guessing:

Combined: 91.209% → 91.163% (-0.046pt), 5 lines out of 10,795 total instrumented lines.

File:line	Explanation	Status
bin/cli_handler.rb:315	--no-app's "Skipped loading Ceedling application" notice — exercised exclusively by the 29 examples migrated out of mixin_ordering_spec.rb; the new integration spec correctly bypasses the CLI entirely	Fixed — new example in dumpconfig_spec.rb
bin/mixin_standardizer.rb:76	Re-keying a config's String '*' matcher-hash key to a Symbol — covered only incidentally via the real temp_sensor project's pre-existing matcher-shaped :defines, never a deliberately-named scenario	Fixed — new example in mixin_loading_spec.rb
configurator.rb:789, configurator_setup.rb:780-781	Configurator's own :plugins-validation failure path — doesn't trace back to this PR's diff (:plugins isn't touched by anything mixin-related here)	Not fixed here — see below

The last 3 lines are more likely a symptom of a pre-existing suite fragility this PR surfaced rather than caused: the original mixin_ordering_spec.rb set ENV['CEEDLING_MIXIN_*'] directly on the real process environment in ~10 places with no scoped cleanup (confirmed — only 2 after hooks in the whole file, neither clearing those vars). That's real cross-example/cross-file environment leakage that can shift which validation branch an unrelated, later-running spec happens to exercise from one CI run to the next, independent of anything this PR changed. Flagging it as a separate, suite-wide cleanup candidate rather than chasing it down further in this PR.

Latest commit (9186564) adds the two fixable cases. New totals: integration spec 17→18 examples, dumpconfig_spec.rb +1 example.