[cailmdaley]

External DR6-only pairing losses now warn and record instead of rejecting a tile. Zero-object tiles also have a valid workflow partition, so the scheduled stages can publish empty products.

Sparse edge tiles 058.220 and 207.238 still exceeded the external pairing budget after #975.
The SExtractor-side wrong-pixels check remains a hard failure.
External-only objects never gain a shape.

• Record for downstream cuts: every newly joined tile's published tile_make_cat.json carries catalogue_match: n_sextractor, n_external, n_paired, n_external_only, n_dropped, external_paired_fraction, and external_warning_budget.
  Join the manifest's unit, converted to float, to final-catalogue TILE_ID.
• Warning threshold: retain max(20, 0.02 N + 50), without the 50 when no pairs exist.
  The external allowance argument/config key is removed; the threshold no longer controls acceptance.
• Empty products: explicit empty stamp shapes, empty EPOCH handling, and safe make-cat reductions support zero-row catalogues.
  The workflow retains its configured chunk count and gives every empty chunk the bounded interval (2, 1), avoiding ngmix's unbounded upper-limit sentinel.
  Ordinary chunk, merge, completeness, and publication paths then receive all expected empty files.
• Decision record: detection.tile_detection states that no pipeline cut applies to external-only losses. Whether the catalogue cuts on external_paired_fraction is open in Decide whether the catalogue cuts on ShapePipe's external_paired_fraction sp_validation#444. The record gives the reason each alternative fails, and pins the removed MATCH_EXTERNAL_UNPAIRED_ALLOWANCE as absent: a config that sets it again fails the decision tests, checked by injecting it.

Nonzero partitions and catalogue measurements are unchanged.
A before/after probe found byte-identical partition documents for 512 nonzero cases.

Tests at c94fbd06: 1153 not-slow and 51 workflow tests pass on a clean checkout in the campaign SIF, inside a four-core Slurm allocation.
The workflow regression resolves the production DAG, executes its partition writer and all 16 chunk readers, and feeds their bounds through the real module-to-publication path.
Restoring zero-object rejection or the unbounded empty range each makes that workflow regression fail.
The matching, durable-record, SExtractor-failure, and empty-module regressions remain covered.
Later commits change only comments and the decision record. The 77 decision-contract tests pass at the head, 9b1d898.

A wholly empty batch still fails the separate external checker's science-selection requirement; this PR does not change that gate policy.

Refs #974.

— OpenAI coding agent on behalf of Cail