test(library): pair every no-fire assertion with a positive control

[AetherAI3]

Part of the vacuous-test sweep across the four repos. Tests only — no rule, IR, or runtime change.

The defect class

A test that only asserts a rule stayed silent cannot distinguish:

• the rule declined to fire (what we meant to prove), from
• the rule can never fire at all (a broken rule, or a frame whose signals don't match)

Both are green. This is the same shape as the vacuous prefetch test fixed in Unlimited-Context-LLM yesterday, which asserted a wall-clock bound on a branch that never executed — green for the wrong reason the entire time.

The repo already had the right pattern

tests/test_examples.py pairs every quiet frame with a firing frame on the same graph, so the graph is proven live before it is proven silent. tests/test_library.py did not, in four places:

Test	Origin
consecutive_loss_circuit	added in #13, an hour earlier
correlation_cluster_guard	added in #13, an hour earlier
atr_volatility_halt	pre-existing
cpi_impulse_pullback_short	pre-existing

The CPI one is the most interesting: the short arm was only ever asserted silent. A short rule that could never fire would have passed unchanged, and the mutual-exclusion claim rested entirely on the long arm. It now gets a mirror hot print and asserts SELL, SELL while the long arm goes quiet.

Result

Each now runs a crossing frame through the same graph and asserts exact intents and timestamps. All four fired — so none of the rules was actually broken. Nothing in the suite could have told us that.

Two of the four were mine, written an hour earlier in the same session that identified the pattern. Worth recording rather than quietly fixing.

401 passed, 2 skipped — unchanged, because these are assertions added to existing tests rather than new test functions.

Sweep status across the other repos

• Unlimited-Context-LLM — clean. test_encode_and_recover guards with assert hits before asserting content; test_engine_beats_baseline_on_reachability is a proper ON/OFF head-to-head; the witness decay test asserts score("doctrine") == 0.9 before checking chatter faded. The one vacuous test there was fixed yesterday.
• aether-agent — clean. The safe-opener refusals assert explicit status and reason before checking nothing spawned, and one carries a positive control (absolute.status === "spawned") in the same body.
• AntiFlock — read-only pass only; no Go toolchain available here to run the suite. The dominant pattern pairs an error check with the count, and empty-assertions usually also pin a specific reason code (e.g. AF-COLLECTOR-FLOW-UNSUPPORTED). No obvious cases found, but this was not verified by execution.