Name every tile_shape store per campaign, not only for image sims

[cailmdaley]

Every tile_shape node-local store is now named per campaign: LOCAL_TAG (a hash of the run dir) is unconditional, not only set for image_sims.
Two concurrent data campaigns over the same tiles, such as A/B arms, can no longer share /local/scratch/sp-<tile> when their same-tile jobs land on one node.
Closes #944, which has the mechanism and our forensics (0 of 895 tile_shape jobs on nibi ever collided).

Changes

• workflow/rules/tile.smk: LOCAL_TAG = sha1(RUN_DIR)[:8] + "-" for every input type; the tile_local() docstring describes the per-campaign naming.
• tests/workflow/test_dag.py: new test_tile_store_is_unique_per_campaign, for both data+psfex and image_sims+fake. Two campaigns at different roots must give each tile different stores, and within a campaign tile_vignets, all tile_ngmix chunks and tile_make_cat must name the same store.
• tests/workflow/params.py: the pin replaces the run-dir hash with <RUN_DIR_SHA1> before hashing, computed from the fixture's run dir rather than read from LOCAL_TAG. The pin stays root-independent, and an empty tag still moves it.
• tests/workflow/params_pin.json: regenerated. Only tile_vignets, tile_ngmix and tile_make_cat change.
• tests/workflow/README.md, profiles/nibi/config.yaml: the normalisation, mutation probes and store-name comment are updated to match.

Resuming a data campaign across this change

The store path is in params.pre, and params is a rerun trigger in both profiles.
So the first sp run of an existing data campaign after this change reschedules tile_vignets, the eight tile_ngmix chunks and tile_make_cat for every finished tile.
As the warning block above LOCAL_TAG in tile.smk explains, that rerun cannot succeed once the tile's exposure stores are reclaimed, and the failed group deletes the finished final_cat.
Running campaigns are unaffected (they use their launch snapshot).
Land this between campaigns. If an existing data campaign must be resumed, run that invocation once with --rerun-triggers mtime code software-env.
Image-sim campaigns are unaffected: their store names do not change.

Tests

Run on nibi in an allocation, in shapepipe-develop-240b37e4.sif with the checkout's src/ on PYTHONPATH:

• tests -m "not slow": 936 passed, 4 skipped, 2 deselected.
• tests/workflow (snakemake 9.23.1, params pins forced with SHAPEPIPE_ON_CANDIDE=1): 19 passed, after --update-params-pin and again without it.
• Mutation probes, each in a disposable copy:
  • Restoring the image_sims-only tag fails test_tile_store_is_unique_per_campaign[data+psfex] and test_unit_pre_changes_at_campaign_boundary (2 failed, 17 passed).
  • Removing the run-dir-hash normalisation fails test_unit_pre_changes_at_campaign_boundary and test_params_pin_ignores_fixture_root (2 failed, 17 passed).

The sweep-ordering hazard noted in #944 is out of scope here.

— Claude (Opus) on behalf of Cail

🤖 Generated with Claude Code

https://claude.ai/code/session_01Dj4HnkNqqPkVCeDeUGEGuz