Report the scientific decisions a PR touches

[cailmdaley]

A PR that changes the logic at a site governed by a decision in astra.yaml passes CI silently when no pinned Values: entry moves, even if the record's rationale is now false. This makes it visible without gating.

• python -m tests.helpers.decisions --diff <base> lists each decision whose @sc-tagged site overlaps the diff (against the merge base; deleted sites are mapped on the old side and marked (base); renames are not changes). Each decision is one bullet saying whether its record changed in the same diff, with the overlapping sites nested under it. Decisions whose record is unchanged come first, since those are the ones where someone should check that the rationale still holds.
• CI appends the list to the job summary on every PR. A failure of the report step writes the error into the summary and never fails the job.
• One sentence in CLAUDE.md points at it.

Tests: tests/unit/test_decisions_diff.py builds throwaway git repos covering edits, deletions, a moved base, and renames/binaries. Each was seen to fail under an injected bug, including a test that unchanged records are listed first.

Example (--diff f477c410~5, trimmed):

- `catalogue_assembly.tile_overlap_handling` — record unchanged; check the rationale still holds
  - `src/shapepipe/modules/make_cat_package/make_cat.py:100-146`
- `shape_measurement.ngmix_seed_mode` — record unchanged; check the rationale still holds
  - `src/shapepipe/modules/ngmix_package/ngmix.py:449-501`
  - `src/shapepipe/modules/ngmix_package/ngmix.py:1241-1440`
- `catalogue_assembly.star_galaxy_classification` — record amended
  - `src/shapepipe/modules/make_cat_package/make_cat.py:143-212` (base)
  - `src/shapepipe/modules/make_cat_runner.py:30-111`

Sites that tag a whole function (e.g. ngmix_runner.py:37-227) flag any edit inside it; tightening those is a record-side change for later.

Claude Opus 5.5 on behalf of Cail

🤖 Generated with Claude Code

[cailmdaley]

This looks good, but i don't think it needs to be a table. i think a set of bullet points would be more productive and easier to read.

[cailmdaley]

Done: the report is now one bullet per decision, with its tagged sites nested underneath. Decisions whose record is unchanged come first, since those are the ones to check. The PR description shows an example, and CI is green.

Claude Opus 5.5 on behalf of Cail