unreferenced_by_name reads 0.25/file for haskell and scheme: a module export list and a type signature are declarations, not references #2823

unreferenced_by_name (the census #2806 renamed from orphaned_logic) reads 0.25 per file for haskell and scheme against a 2.50 corpus median — 12 of each language's 13 probe functions read as referenced when nothing calls any of them. This is #2806's defect with the sign flipped: there the census counted a function as unreferenced that the language reaches without naming it; here it counts a function as referenced on a name occurrence that is a declaration, not a reference.

Measured 2026-09-06 on engine fix/2806-orphan-invocation-model (branched from main 87cfc933) / corpus main 5cf94e7, per-file census over keyword-rosetta/data.

The two forms

_is_orphan (detector.py) counts occurrences of the name outside [start_idx, end_idx), discounts one of them when the span contains none (the declaration, which for several slicer modes falls outside its own span), and discounts the offsets captured by the registry's _visibility_export rule (#2774). That budget of one implicit + one declared form is exactly enough for tcl/shell/ruby/powershell/assembly. haskell and scheme each write the name a third time:

module A (probeGlobals, probeTest, probeSafety) where   -- export list: one occurrence per name
probeGlobals :: Int -> Int                              -- type signature: another
probeGlobals env = env                                  -- the equation (span starts after the name)

(export probe-globals)          ;; scheme declares no _visibility_export, so this reads as a use
(define (probe-globals env) env)

What is and is not enough

• scheme needs only the _visibility_export declaration it never got. Adding ^\([ \t]*export[ \t]+([a-zA-Z][\w\-!?*/<>=+]*) moves the corpus 0.25 → 3.00 per file, not 2.50 — because main.scm's entry does not call its probes either, so scheme also carries the corpus plant gap orphaned_logic reports 100% dead code for five languages whose invocation model never names the callee — the family #2727 closed without covering #2806 found in abap/m4/makefile/objective-c, hidden until now behind the export false-negative. Both halves have to land together or scheme swaps an amber cell for a worse one.
• haskell cannot be expressed with the mechanism as it stands. export_name_starts is built as frozenset(m.start(1) for m in _vis_export.finditer(...)) — one capture group, one name per match — and a Haskell module export list holds an arbitrary number of names inside one pair of parens. Declaring the type signature instead (^([a-z]\w*)[ \t]*::) is a legal single-capture rule but does not move the count (measured: still 0.25), because the export list alone is enough to clear every flag.

So the fix is (1) let _visibility_export yield more than one name per match — every group, or a second _visibility_export_list whose capture is the whole list and whose names are split by the engine — and (2) declare it for haskell and scheme, and (3) plant the missing dispatch calls in keyword-rosetta/data/{haskell,scheme}/main.* in the same wave.

Why it matters

The metric feeds risk_tech_debt and the orphan→api conversion, so an under-reading language looks maximally clean on both. It is also the last shape of the census that is still measurement rather than code: after #2806, orphan-detection-is-name-recurrence's remaining languages are these two plus the no-functions family (html, markdown, yaml, sqlite, dockerfile, css), which is a different question (#2549's undefined family).

Part of #2812. Successor to #2774 and the sign-flip of #2806; related #2727, #2754.