Bind stores and receipts to sources and harden cohort execution

[CraigMyles]

This keeps v0.1 local output names stable, while preventing a same-named slide from silently reusing or appending to a store belonging to another source.

Changes

• Bind receipt skips to the slide ID, credential-free canonical source, expected output target, and live store provenance.
• Check the root and every grid before non-force store reuse or append, failing clearly on missing, malformed, inconsistent, or mismatched source provenance.
• Make --force bypass receipt fast paths and deliberately rebuild the target.
• Reject duplicate derived output IDs across the complete embed-many cohort before sorting, sharding, or loading models, including repeated identical manifest rows.
• Propagate parallel worker model-load failures as a non-zero command exit without counting worker failures as failed slides.
• Replace receipts atomically, treat malformed JSON as incomplete, preserve existing file permissions, and avoid overlong temporary filenames.
• Add target-aware verify --out-dir handling and update the shipped SLURM array accordingly.

Compatibility notes

• Ordinary local slide IDs remain basename-based, so existing v0.1 output names do not change.
• The three-argument is_complete(...) call remains valid but now fails closed when no current source is supplied.
• Per-output locking across independent processes remains deferred; the cohort preflight prevents duplicate writers within one discovered cohort.

Validation

On Spark:

• uv run ruff check .
• Focused regression suite: 116 passed
• Full non-slow, non-network suite: 407 passed, 10 skipped, 87 deselected

Refs #6