Range the replay-safety guard over the substrates the factory can build

[xmap]

What was wrong

The guard meant to keep decider_replayability in lockstep with the deciders had no independent side. It asserted a hand-written set of four refs against the classifier's own hand-written sets, so it agreed by construction, and it never looked at an adapter at all.

It therefore reported nothing while the llm substrate shipped a model_ref of provider:model, a value no literal set could ever contain. Asking is_replay_safe about an LLM-steered run raises rather than answering. No production caller consults it yet, so the defect is dormant, but the module docstring claimed a coverage it did not have.

The fix

Key the classification by DecideSubstrate instead. That Literal is the independent side: decide_port_config maintains it so its factory knows what it can build, a substrate missing from it cannot be constructed, and whoever adds an adapter must edit it for reasons that have nothing to do with replay. The test now asserts set equality in both directions, so a seventh substrate fails CI until classified and a deleted one fails until its classification goes too.

Two substrates do not record their own name, and the classifier now says so rather than pretending otherwise:

• staged returns its child's advice unchanged, so its own name on an iteration means the composite grew a ref it should not have. That input raises.
• llm records provider:model, an open set rather than a constant, recognised by shape.

Verification

Verified by mutation: adding an unclassified substrate to the Literal, misclassifying llm, disabling the llm route, and removing the composite refusal each fail exactly their own tests and nothing else. One test was dropped along the way for failing that check, having compared a result against the same map it came from.

14,934 unit / 37,424 architecture passing, full-repo pyright clean.

Naming

Replayability / SUBSTRATE_REPLAYABILITY / replayability_of, as StrEnum. The naming reviewer overturned an earlier ReplayClass: there are zero pre-existing *Class enums, and 114 of 116 enums in the tree are StrEnum.

🤖 Generated with Claude Code

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
apps/api/src/cora/operation/adapters
decider_replayability.py
Project Total

This report was generated by python-coverage-comment-action