fix(core-engine): an export construct that names many functions is still a declaration (#2823, roadmap Phase 3)

[squid-protocol]

Closes #2823. Part of #2812.

unreferenced_by_name read 0.25 per keyword-rosetta file for haskell and scheme against a 2.50 corpus median — 12 of each language's 13 probe functions reading as referenced while nothing called any of them. That is #2806's defect with the sign flipped: there the census called a function unreferenced that the language reaches without naming it; here it called a function referenced on an occurrence that is a declaration, which corollary 2 of the stated contract says must not clear the flag.

One outside occurrence each, and it is the export

module A (probeGlobals, probeTest, probeSafety) where   -- outside the span
probeGlobals :: Int -> Int                              -- INSIDE the span
probeGlobals env = env                                  -- INSIDE the span

(export probe-globals)                                  -- outside the span
(define (probe-globals env) env)

The Haskell type signature needed no rule of its own: haskell's slicer span starts at the signature, so that repetition is already inside the function's own definition. That explains the issue's own unexplained measurement — declaring ^([a-z]\w*)[ \t]*:: as an export form moves the count not at all.

_visibility_export_list

#2774's _visibility_export captures one name per match, which is the whole statement for export -f foo and its four siblings and cannot express a construct that names every exported function at once. This adds the plural form: its capture groups are regions holding exported names, and _export_declaration_offsets records the start offset of every name token inside a region.

Two keys rather than one generalised key, deliberately. The singular form records its group's own start offset verbatim, and assembly exports .foo while ruby exports save! — whose first and last characters a region tokenizer does not treat as part of a name. Tokenizing those captures would put the offset one character off and silently stop discounting them. A language declares one key or the other, and test_the_two_export_keys_never_disagree_about_a_language is the gate on that.

haskell declares the module header's parenthesised list (bounded by the ) where anchor, since [^)]* stops at the first Opt(..), and every one of pandoc's seven headers terminates exactly right); scheme declares both of its clause forms, R7RS (export a b c) and Guile #:export (a b c).

Measured, both corpora

corpus	before	after
keyword-rosetta, haskell	0.25 /file	3.00 /file
keyword-rosetta, scheme	0.25 /file	3.00 /file
keyword-rosetta, other 44 languages	—	unchanged
language-crucible, haskell	2 of 147 units (1.4%)	41 of 147 (27.9%)
language-crucible, scheme	0	0 — neither file uses an export clause outside a macro template

No other language can move: the singular path is byte-identical and only these two declare the new key.

3.00, not the 2.50 median, because both languages also carry the same plant gap #2806 found in abap/m4/makefile/objective-c — hidden until now behind the export false-negative — where main's entry dispatched one of its three probes. The corpus PR plants the other two and both land on 2.50.

The crucible's 39 new hits are correct. Every one is a pandoc utility whose name occurs exactly three times in its own file — export list, type signature, equation — with no caller there. That is corollary 5 working as designed: Text.Pandoc.Shared is a library module of 100 exported helpers, 36 of which nothing in that file calls, and saying so is the input galaxyscope.py's orphan → api conversion needs. Same shape as #2774's shell precedent (0 → 3 per file).

The bless

Nine value differences, identical in both fixtures, every one of them haskell/pandoc:

what	was	is
Shared.hs Unreferenced By Name	0	36
Filter.hs Unreferenced By Name	2	4
Options.hs Unreferenced By Name	0	1
those three files' Tech Debt Exposure	0.0% / 41.35% / 0.0%	97.04% / 85.83% / 9.7%
haskell/pandoc avg tech debt	3.76	17.51
global avg_tech_debt	27.332	27.386

The tech-debt rows are the census's own weight arriving downstream (slop_stress counts each unit at 2.0). No topological X/Y/Z coordinates moved at all — unusual for a bless, and the reason is that no node's structural mass changed.

One neighbour moved — a pre-existing defect this made visible

With scheme finally producing orphans, its api_orphan_credit went 0 → 3 in a/b/c: the double count #2731/#2734 exists to prevent. api_declared_orphans tokenises the api-matched line with \b\w+\b, which cannot produce a token holding a hyphen, so probe-globals is never recognised in (export probe-globals).

haskell is the control that proves the diagnosis — its probe names are all-\w, its api_declared_orphans reads 3, and its credit correctly stays 0.

Not new, and much larger than the control corpus suggests: cobol carries it on main today, and over the crucible api_declared_orphans moves 4,983 → 5,122 under a name-aware test, +139 across zig (@"..." quoted identifiers), lua (bit.bor, M.start) and powershell — whose count is 0 for the entire language, because a Verb-Noun name always contains a hyphen. Filed as #2827 rather than fixed here: it moves five languages and needs its own bless.

Docs

• docs/unreferenced_by_name_contract.md: corollary 2 gains the plural form, the audit table's haskell/scheme row moves 0.25 → 2.50, and a new section states the whole finding — why two keys, what the crucible says, and what the bless moved.
• how_to_add_a_language.md rule 19 gains the opposite-mistake bullet: a census that reads too clean is the same defect as one that reads too dirty.

Verification

• Full suite: 7,914 passed, 4 skipped, 9 xfailed.
• crucible_check.py: both fixtures re-blessed, diff scoped above (9 values each, no unattributed mover).
• rosetta_audit.py --baseline-bin <engine main>: 46 languages, 0 regressions.
• tri_comparison_chart.py --all --ci: 3 languages, all OK.
• audit_check.py (ruff / mypy / dead-key / ast-accuracy): all clear.

Cross-repo

• Corpus: plant: main dispatches its probes in haskell/scheme, and scheme owes a credit re-bless (gitgalaxy#2823) keyword-rosetta#83 (branch plant/2823-haskell-scheme-dispatch) — plants the two dispatch calls, re-blesses scheme's api_orphan_credit cell against a new api_declared_orphans tokenizes api lines with \b\w+\b, so a hyphenated or qualified name is never recognised as already-declared (#2731's double count returns for cobol and scheme) #2827 ledger entry, and narrows orphan-detection-is-name-recurrence from 8 languages to 6.
• This merges first; that one goes green against gitgalaxy main once it lands.
• Follow-on: api_declared_orphans tokenizes api lines with \b\w+\b, so a hyphenated or qualified name is never recognised as already-declared (#2731's double count returns for cobol and scheme) #2827 returns scheme's and cobol's api_orphan_credit cells to 0 and owes another corpus re-bless.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab

[squid-protocol]

rosetta-audit failure: the companion re-bless, labelled

REG  scheme             FAIL scheme: 3 mismatch(es)
       - a.scm: api_orphan_credit expected 0, got 3
       - b.scm: api_orphan_credit expected 0, got 3
       - c.scm: api_orphan_credit expected 0, got 3
rosetta_audit: 46 language(s) checked -- 1 regression(s), 0 pre-existing, 0 broken.

This is the exact mirror of squid-protocol/keyword-rosetta#83's own red CI, and the two together are the whole cross-repo state: rosetta-audit checks out the corpus at main, whose scheme manifest still expects api_orphan_credit: 0 — the value it reads when (export probe-globals) is still wrongly clearing the orphan flag. With this build there are orphans again, galaxyscope.py's Contextual Baseline Fix converts them, and the cell reads 3.

The corpus PR blesses those three cells to 3 against a new ledger entry (api-declared-orphans-blind-to-non-word-names, #2827) and cannot go green until this merges; this cannot go green until that one does. Applied rosetta:rebless-owed — the label the workflow documents for exactly this shape.

Worth stating explicitly: only scheme moves, and only on this one downstream cell. haskell passes here, on corpus main, with no plant — its api_orphan_credit is 0 in both engine states, for two different reasons (no orphans before; all three suppressed by api_declared_orphans after). The census itself is not gated in any manifest, so haskell's and scheme's 0.25 → 3.00 move is invisible to this audit; what it caught is the single consequence two layers downstream, which is #2827.