[10-3] Re-run the Goal 6/7 grids where the representation verdict could flip

[duckyquang]

Goal 10, Task 3 — [10-3] · GEML Goals 6–12 (roadmap, README §5) · Size: L

Independence

Hard prerequisites (genuinely required): [10-2], [6-5], [7-3]. These cannot be stubbed: it re-runs the merged Goal 6/7 grid machinery over v2-derived datasets, targeting only the cells the [10-2] deltas say could flip.

What can start early, before the prerequisites merge: claim the issue, write the config and the tiny-fixture smoke test — only the work that touches real merged code or production artifacts must wait.

• Exclusive write set — no other issue edits your owned paths, so your PR cannot conflict with parallel work.
• Branch: issue/10-3-<short-slug>.

Objective

Selectively re-run the Goal 6 equivalence and Goal 7 step grids on corpus v2 where the [10-2] compression deltas indicate the representation verdict could change; update the record either way.

Exclusive file ownership

Owned paths — this branch may edit only these paths:

• src/geml/experiments/goal10/rerun_grids.py
• configs/goal10_grids.yaml
• docs/goals/goal10/GRID_RERUN.md
• tests/experiments/test_goal10_grids_smoke.py

Read-only dependencies:

• v2 studies [10-2]
• grid machinery [6-5]/[7-3]
• pair/step builders [6-1]/[7-0]

Tasks

• Identify flip-risk cells from [10-2] deltas; document the selection rule before running.
• Rebuild v2 pair/step data with the existing builders (train-only, leakage-free).
• Re-run the selected cells with identical configs and 3 seeds; compare against v1 results.
• State explicitly whether any verdict flipped, narrowed, or held.
• Keep every cell row, including unchanged ones.

Acceptance criteria

• The cell-selection rule is pre-registered in the config, not post-hoc.
• v1 and v2 results are compared like-for-like.
• Verdict changes (or none) are stated explicitly.

Validation

• python -m pytest tests/experiments/test_goal10_grids_smoke.py
• Run the production command in configs/goal10_grids.yaml.

Out of scope

• Do not expand the grammar further or touch v1 artifacts.

PR checklist

• I edited only the owned paths listed above.
• I did not inspect or reuse the previous GEML implementation.
• Tests use temporary/tiny fixtures and do not require production outputs.
• Failures and unsupported cases are retained and reported, not silently dropped.
• I included exact reproduction/validation commands in the PR description.
• I documented any scientific metric or assumption change.

Clean-room rule: Implement only from the current repository specifications and authoritative public sources. Use the v0 prototype only to understand what was measured and why — never as a code template. If an interface is ambiguous, stop and ask rather than modifying another issue's contract.