Deterministic helper for the tunex rule-gen flow (tunex docs/10).

mix credence.fix_tests test/pattern/foo_fix_test.exs [more_test.exs ...]

An LLM building a rule writes assert fix(Rule, input) == expected but can't byte-predict the rule's exact output, so expected is wrong → the fix test fails → the Gate rejects the (often correct) rule, and the agent burns turns guessing. This task removes that class WITHOUT the AI: for each input = """…""" / expected = """…""" / assert fix(Rule, input) == expected block in a Pattern rule's fix test, it RUNS the rule on input and rewrites expected to the actual output — which is exactly the project's own convention ("expected is the rule's REAL output — run it, copy the string").

Safe + conservative:

• Only Pattern rules (it calls fix_patches/2); the module is derived from the test path and must be loaded (so a rule that doesn't compile is skipped — its failure is a real bug for the agent, not ours).
• Only the fix(Alias, input) == expected shape with resolvable heredoc operands; anything else is left untouched.
• NO-OP fixes (actual == input) are NOT rewritten — that would create a expected == input test that fails the "real transformation" meta-gate; a no-op fix is a genuine rule bug, left for the agent.
• Correctness is still guarded by the rule's equivalence test, not by this (auto-captured) expected.