test(ADR-012): assert the lr_/by_ target prefixes — the guard the ADR asked for and never got

[Polichinel]

ADR-012 proposed this guard itself — "Existing tests (test_config_completeness.py) can be extended to assert that all regression_targets use the lr_ prefix and all classification_targets use the by_ prefix" — and it was never written. Nothing checked either prefix, for about three months.

Same family as #367: a targets declaration that no test inspected. #374 closed the metric half by loading every config through pipeline-core's CoreConfigSniffer. The sniffer validates the targets↔metrics pairing and never the target names, so this closes the other half.

What the tree already looks like

Measured across every non-fixture source:

	count	prefixes
regression_targets	190	all lr_
classification_targets	24	all by_

The convention already holds everywhere. This pins it; it does not repair anything.

Fixtures

Nine fixture entities declare a deliberate synth_target. They are excluded via meta/fixtures.json, loaded rather than hardcoded — the rule tools/partitions/fileops.py and tools/catalogs/create_catalogs.py follow, pinned by TestFixtureSetConsistency (C-61, #99).

That exclusion is itself pinned by a test, so it stays a declared fact rather than a circular one: a real model adopting synth_ still fails.

Placement

Own file, not an addition to test_config_completeness.py. That file asks whether required keys are present; this asks whether the values are well-formed. Different question, different file.

Both ADRs updated to say the guards exist

• ADR-012 — implementation note, with the three-month gap stated rather than glossed.
• ADR-009 — the conventions table gains a targets↔metrics row. Worth noting the gap was wider than the one missing key: that table previously mentioned no targets or metrics keys at all. The row attributes the rule to pipeline-core's sniffer rather than restating it, so the two cannot drift.

Verified it bites

Injecting 'ln_sb_best' into bright_starship turns it red, naming the model and the offending target; removing it turns it green.

My first attempt at that check used "by_sb_best" while the file uses 'by_sb_best', so the injection silently changed nothing and the "test" passed. The guard was fine — the verification was the vacuous thing. Worth recording, because a check that cannot fail looks exactly like a check that passed.

Verification

• ruff check . clean
• Full suite: 7753 passed, 219 skipped, 6 xfailed, 1 xpassed, 0 failed