PLT-464: regression fence — profiles parse + legacy default path

[bdchatham]

Implements PLT-464 (M1.6) — a CI fence for the additive-by-construction promise.

What

A config-package test (config/regression_fence_test.go, run by make test/CI — no manual step):

• Globs every profiles/*.json and asserts each resolves through the binary's real settings path (LoadSettings→ResolveSettings→Validate, the viper merge where additive defaults apply) — strictly stronger than the existing raw-unmarshal profiles_test.go.
• Asserts a no-new-fields config resolves to the legacy default: ArrivalModelClosedLoop, distributions nil — the resolution that gates main's openLoop branch. Asserted against DefaultSettings()/the constant, not literals.

All 8 shipped profiles parse + resolve; a falsification probe confirmed the legacy assertion has teeth (an open_loop config survives as ArrivalModelOpenLoop).

Verify

make verify green (lint-pin + lint 0 issues + test + build + --help + bindings in sync).

🤖 Generated with Claude Code

[cursor]

PR Summary

Low Risk
Test-only changes with no production config or runtime behavior modified; risk is limited to CI failing if additive defaults or profiles drift.

Overview
Adds PLT-464 CI regression tests in config/regression_fence_test.go so the additive-by-construction config contract stays true: shipped profiles still parse and resolve like the binary, and configs without new workload fields still land on the legacy closed-loop path.

Every profiles/*.json is discovered via glob and run through unmarshal → LoadSettings → ResolveSettings → Validate, which is stricter than raw JSON unmarshal alone. Profiles that omit additive knobs must resolve to ArrivalModelClosedLoop with nil key/size distributions; profiles that opt into arrivalModel, maxInFlight, or distributions are skipped for that legacy assertion. A minimal synthetic config pins the same defaults independent of which profiles ship.

registerSettingsFlags is extracted as the shared cobra flag set for viper tests; TestArgumentPrecedence in settings_test.go now uses it instead of duplicating flag declarations.

^{Reviewed by Cursor Bugbot for commit 95fa12a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3374877. Configure here.}

[cursor]

Legacy fence skips explicit defaults

Medium Severity

usesNewFields treats any present arrivalModel string or maxInFlight JSON key as opting in, including values that only restate legacy defaults (closed_loop, default max-in-flight). Those profiles skip TestProfilesNoNewFieldsResolveToLegacyPath, so the CI fence no longer checks arrival model or nil distributions for them and can hit checked == 0 if every profile documents defaults explicitly.

Additional Locations (1)

• config/regression_fence_test.go#L117-L121

 

^{Reviewed by Cursor Bugbot for commit 3374877. Configure here.}