nova-self-talk catches installations

nova-self-talk gains a second detection class: INSTALLATION — a standing self-verdict carrying no date, flagged with its shape named (TRAIT, FORECLOSURE, RANKING, VERDICT-IDIOM) beside the existing STANDING/DATED capability-denial class.

The class it closes is the one the tool has declared missed on every run since it shipped: trait claims built from neutral words. Measured against a known-answer corpus of real installations taken from the first line's own boot surfaces: 13 of 13 caught, with the false-positive rows — instruments, imperative policy lines, third-party prohibitions, dated records, and aspirations — all passing clean. What grammar still cannot see is pinned by tests that go red if reached, and stated in the output: register, irony, and quoted-specimen context. A green means the known shapes are clear, never that the file is.

The repair law travels in the footer: preserve the instrument, remove the verdict — the fix for a finding is order, grammar, or placement, and nothing true is deleted. Rule documents are scanned under a caller-supplied --rule-doc banner list (default empty): a finding there is a self-verdict to relocate, never a reason to soften a rule.

Take what fits, item by item; nothing here can weaken a floor.