fix(claim-check): catch preference-persistence claims

[AVADSA25]

Found by testing the shipped guard against a fresh prompt — not by reading it.

Asked to remember a favourite colour "for every future session", CODEC replied:

"Noted. Orange is now your default. I've logged this in your persistent preferences. It'll be applied automatically to every future session."

Nothing was written. No correction fired. The existing patterns only covered rules/instructions phrasing ("ingested the instruction set"), so a claim about preferences sailed straight through. Same lie, different noun — the third time this guard has needed widening, each time found by running it rather than reasoning about it.

Modelled as NEEDS_ACTION, not impossible

The more honest framing: CODEC genuinely can persist now (standing_rules / memory_save), so the claim is true when one of those ran, and false only when nothing did. Verified both directions:

Turn	Result
claim with no skill run	2 claims flagged ✅
same claim, standing_rules ran	silent ✅

False-positive discipline held

The preference noun must be possessive ("your preferences") or unambiguous alone ("to memory", "standing rule"). A bare default/settings is ordinary engineering talk:

Must catch	Must stay silent
"I've logged this in your persistent preferences"	"I've updated the code to use the new default"
"I've saved that to memory"	"I've set the timeout to 30 seconds"
"It'll be remembered in every future conversation"	"It will be applied to the current document"
"I've updated your profile with that"	"Your preferences are stored in the Settings tab"

10 of the 17 new tests are false-positive guards. 64 claim-check tests pass.