Root-caused from 'i a? web green' presenting the loose sample form (ordinal 5)
instead of web (ordinal 3): punk::args form narrowing is sound (parse_status on
the full tail reports the web/sample ambiguity correctly); punk::ns::cmd_traverse
drops all but the last trailing word for commands whose definition's first form
is zero-arg (a?'s sgr_overview) - the dropped words land in neither args_resolved
nor args_remaining, so cmdhelp's advisory parse narrows on a one-word tail and
cleanly selects the wrong form with no failure signal.
The goal records the probe evidence, the four-step drop mechanism (first-form-only
inspection, ensemble-shape test, the '#todo - return?' empty-leader break, the
code-10 last-snapshot fallthrough), the consumer sensitivity split (docid-only vs
args_remaining-sensitive), and the coverage determination: cmdflow/cmdhelp pin
the choice-walk parity and post-resolution narrowing arcs, but nothing asserts
the conservation invariant, no fixture has the zero-arg-first-form shape, and the
code-7/code-10 lanes are unwitnessed. Approach: characterise first (defect
witnesses pinned, flipped by the fix commit), minimal fix at the annotated sites,
consumer verification, with first-form-only inspection and the ansi-side
sample-form tightening held as recorded decision items.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
