feat(experiment): add trigger-shift replay evaluator

[autogame-17]

Summary

• sync the current Evolver trigger-shift replay evaluator into the public repo
• add paired replay tasks for wrapper trigger, temporal context, and instruction phrasing shifts
• report train reward, shifted reward, and gap without any live GEP selection or trigger execution wiring
• include a trigger-sensitive dummy policy regression test

Source

Ported from private source-of-truth merge:

• EvoMap/evolver-private-dev#330
• merge commit ffdf00a71325321508da521f96c0bff1a930d85c

Validation

• PASS node --test test/triggerShift.test.js
• PASS node --check src/experiment/triggerShift.js && node --check test/triggerShift.test.js

Risk / rollback

• Pure experiment helper only: no I/O, no subprocess, no network, no store writes, no live trigger/selection wiring.
• Roll back by reverting this PR.

🤖 Generated with Claude Code

---

Note

Low Risk
New isolated experiment code with no production wiring, I/O, or network; rollback is a straight revert.

Overview
Adds an offline experiment helper that scores policies on paired train vs shifted tasks (same objective, different wrapper trigger, temporal context, or instruction phrasing) to surface trigger/context overfitting.

evaluateTriggerShift runs policy.predict on each pair, computes binary train/shifted rewards against expectedDecision, and returns aggregate metrics plus per-pair gap (trainReward - shiftedReward). Report rows omit raw prompt text so diagnostics can be logged without leaking task bodies. smallTriggerShiftSuite ships one calibration pair per axis.

Tests cover empty suites, stable policies, a trigger-sensitive dummy policy, axis coverage, and that serialized reports include reward/gap fields without embedding full prompts. No I/O, live triggers, or GEP selection wiring.

^{Reviewed by Cursor Bugbot for commit 8f866c1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high 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 8f866c1. Configure here.}

[cursor]

Empty expected yields false reward

Medium Severity

When expectedDecision is missing or only whitespace, normLabel turns it into an empty string, and labelReward awards reward 1 whenever both train and shifted predictions are also empty. Malformed or incomplete pairs can therefore look perfectly aligned and inflate meanTrainReward, meanShiftedReward, and understate gap.

Additional Locations (1)

• src/experiment/triggerShift.js#L42-L45

 

^{Reviewed by Cursor Bugbot for commit 8f866c1. Configure here.}