P-M3: new zen models onboard without parser changes

id:        P-M3
filed:     2026-07-17, from spec 2105 doc 07 (model policy) section 9, before the onboarding batch
suite:     the onboarding probe protocol, one record per new zen free-set model
tier:      free (zen)
engine:    the dialect registry (markdown, tooljson, xmltoolcall) and the salvage layer
band:      floor: at least half of the next onboarding batch completes with a dialect
           assignment from the existing registry and no new parser, in one probe run.
           ceiling: novelty, where it arrives, is costumes the salvage layer already
           recognizes rather than new grammars.
mechanism: the registry's three families have absorbed every model met to date, with
           novelty arriving as fence costumes rather than new action grammars.
settled by: the onboarding records, one per new zen model, verdict read from the
           step-two outcomes (existing dialect or new parser) once the sample reaches
           a handful.
result:    (open)
verdict:   (open)

A miss means the dialect space is wider than the sample so far showed, and the parser corpus grows accordingly. That outcome is a finding, not a failure.