fix(core-engine): the orphan census counts references, not name repetitions (#2727, #2728)

[squid-protocol]

Closes #2727.
Closes #2728.
Part of #2669 (F.4).

orphaned_logic was decided by one line — token_counts[name] <= 1 over the whole code stream —
so whether a function read as orphaned depended on how many times the language's syntax writes its
name, not on whether anything calls it. Three distinct defects fall out of that, two filed and one
found while fixing them.

1. #2727 — syntax that repeats the name masked every orphan

Ada closes with end Probe_Globals;, LiveCode ends a handler by name, Haskell writes the type
signature above the equation. The count is already 2 before anything calls the function, so those
languages could never report one: livecode found 3 orphans across the crucible where the
scoped test finds 127; ada and livecode read 0 per file in keyword-rosetta against a 2.50
baseline.

The fix shape as filed was unsafe, and I corrected it. "Count outside [start_idx, end_idx)"
assumes the span contains the function's declaration. It often doesn't:

• shell — Mode-B spans start at the {, so a K&R make_module() on the previous line sits
  outside its own body (curl/initscript.sh);
• ruby — spans are offset outright; 66 of 132 sampled functions have the declaration outside.

Counting naively there makes the declaration itself read as a reference and clears the flag —
measured: ruby 51 → 44 orphans, html 5 → 2, shell 8 → 3, all real ones lost. So when the span
holds no occurrence of the name, the definition site is outside it by construction and exactly one
outside occurrence is discounted. With the correction, no language decreases for that reason.

I also tried span ∪ start_line; it fixes ruby and html but not shell, because start_line points
at the { rather than the line carrying the name — a separate metadata observation worth its own
issue.

The Haskell caveat the issue raised is moot: the span does start at the type signature
(probeGlobals's span begins at probeGlobals :: Int -> Int). Haskell still reads ~0 orphans, but
for a reason the issue didn't anticipate — its module export list names every probe outside every
span. That's keyword-rosetta's already-ledgered api-export-names-suppress-orphan-credit shape,
corpus-side, not an engine gap. The one orphan it does find (entry, unexported and uncalled) is
correct.

2. A third defect neither issue named

\b\w+\b cannot produce a token containing -, : or .. Any name holding one therefore counted
zero and satisfied <= 1 unconditionally — it was an orphan however many times it was called:

language	before	after	name shape
cpp	1054	720	Class::method
tcl	210	148	::ns::proc
powershell	56	3	Verb-Noun
scheme	60	10	hyphenated define
abap	41	8	hyphenated names
cobol	52	45	hyphenated paragraphs

This also corrects part of #2727's own analysis. Its third table row — abap, cobol, jcl, m4,
makefile, objective-c, scheme at 3.25 orphans/file — is attributed there to names "written once,
invoked by other means". For the hyphenated members the cause was tokenization: the counter
could not represent the name at all. scheme goes 3.25 → 0.25 orphans/file in the corpus on that
alone.

3. #2728 — igniter-keyword buckets entered the checks as ordinary functions

dockerfile's four same-bodied RUN slices scored state_slop_duplicates 1 — the only nonzero
duplicate cell across all 46 corpus languages, and a phantom. css's keyframes slices cleared each
other's orphan flag by repeating the keyword.

The bucket set is derived from each language's own func_start, never hand-listed, so it cannot
drift: _closed_literal_capture() returns names only when the capture is provably a fixed literal
alternation and the empty set otherwise. Exactly three languages qualify — css, dockerfile,
html — and test_closed_literal_capture_is_generated_not_hand_listed pins that membership, so a
new keyword in any of those alternations is picked up automatically and a rule that can produce a
real identifier can never opt in by accident.

Deliberately not added to _UNCOUNTABLE_SLICE_NAMES, as the issue suggested. That set governs
the function population (#2691); removing css's ~150 at-rule slices would take functions_found
to 0 and make every per-function descriptor undefined — #2689's markdown/html shape. That is a real
question, owned by keyword-rosetta's slicer-segments-statements-not-functions entry, and it is not
#2728's measured harm. Recorded in the code rather than silently skipped.

Verification

• Full suite 7632 passed serially (CI's mode); 5 new detector tests, one per behaviour above.
• tree_sitter_accuracy_audit --ci --all 30/30 OK; tri_comparison_chart --all --ci 3/3 OK;
  audit_check.py clear; ruff and mypy at baseline.

Golden masters re-blessed: 2278 / 2277 differences. The scoping proof is what moved inside
7. Structural Signatures: Orphaned Logic 581 + Public Exports 71 = 652 of 652. No other
structural signal moved at all, which is exactly what a change confined to the orphan census should
look like. Everything else is that signal's known consumers:

class	count
Tech Debt Exposure / tech_debt (orphans are dead weight)	649 + 101
Topological Coordinates X/Y/Z (corpus-wide mass re-solve)	489
API Exposure / Documentation Exposure (via the Contextual Baseline Fix)	98 + 82
Structural Magnitude / Public Exports	71 + 71
group + ecosystem + forensic roll-ups	388

Net crucible orphans 7397 → 7803 (+5.5%): increases where the span fix removes a mask
(typescript +333, livecode +124, php +100, jcl +50, go/rust +46), decreases where the
tokenization fix stops manufacturing them.

Cross-repo

Companion: squid-protocol/keyword-rosetta re-bless, owed after this merges — carries the
rosetta:rebless-owed label per keyword-rosetta/docs/GATING.md's cross-repo flow.

Corpus impact is unusually small for a change this size: 4 cells, api_orphan_credit on
dockerfile/a.dockerfile (1 → 0) and scheme/{a,b,c}.scm (3 → 0). The rosetta manifests do not
gate orphan counts directly — they see them only through the orphan→api conversion — so the
languages whose orphans move but whose api rule already declares them (ada, livecode, haskell)
record no manifest change at all.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WfeHRWra1d6Z5uSReFajSq

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab