The edit witness is h plus six hex, so it lexes as a string

The exarch line-hash witness is the letter h followed by six hex characters, because a bare witness must round-trip through ral’s value model as a String, and an all-hex token can be all digits. Bare-word elaboration classifies a token syntactically before the type checker runs: Val::from_word sends any word that parses as i64 to Val::Int and only the rest to Val::String (surface-syntax). An agent reads a witness from view and types it back as a bare edit argument, so an all-digit digest like 152347 elaborates to Val::Int(152347) while the hash edit recomputes is Val::String("152347"). equal is structural with no Int×String arm, so the endpoint check rejects a correct witness and the agent loops re-issuing the same edit. The h prefix removes the ambiguity at the source: the witness can never parse as a number, so it is a String at every call site and the comparison compares like with like.

Alternatives, and why they lose:

• String-coerce the comparison in edit (to-string $start-hash). Incomplete, not merely inelegant: a leading-zero digest is already Val::Int(12345) by the time edit runs — the zero is gone at elaboration — so to-string yields "12345" ≠ "012345". The string-ness must be preserved before classification, not reconstructed after. Correctness rules this out, not taste.
• An Int×String arm in equal. A wrong repo-wide semantic — equal 1 "1" would become true. The fault is the witnessed-edit layer reading a string token as a number, not equality.
• Type-directed literal elaboration — let a numeric-shaped word stay a String when it lands in a String-typed hole. The principled language-level fix and not a coercion at all (bidirectional elaboration, fromInteger-style). But elaboration is syntax-directed and precedes typing: elab_expr calls Val::from_word with no expected type in scope. Achieving it needs bidirectional elaboration threading expected types to leaves, or a polymorphic literal with defaulting rules — disproportionate to a witness format. Left as an open language question, not taken here.
• Drop bare-word Int classification (write !{int 50}). Relocates the coercion to every numeric builtin boundary as an implicit String→Int, and ral is a shell where bare integers are the common case. Strictly worse on the very axis it was meant to clean up.

The change is the witness format, owned by exarch: the producer (view, grep-files) and the consumer (edit) both call line_hash, so the prefix is internally self-consistent; only the documented six-hex contract moves with it.

See also hash-addressed-editing, builtins, shell-eval.