fix: stop table-driven tests from being reported as clones

[machado144]

Closes #31

Table-driven tests matched their own rows. The reported "instances" were overlapping sliding windows over a single composite literal, not duplicated logic — which contradicted the README's promise that data literals are excluded.

Two defects

1. Overlapping instances. Seeds inside a file are spaced minTokens apart, but greedy extension grows every block to totalTokens, so neighbouring instances overlapped — one stretch of text counted several times.

#1  type-2  24 lines  3 instances
      table_test.go:15-38
      table_test.go:33-56   <- overlaps 15-38
      table_test.go:51-74   <- overlaps 33-56

2. Data tables. The rows of a collection literal are structurally identical by design. Literals are now recorded as token spans, and a clone whose every instance lies inside one span is dropped as data.

On the approach

Excluding literal tokens from detection (the issue's suggestion 1) was implemented first, then rejected on measurement. It punches holes: a short inline []TokenizedFile{a, b} in the middle of a duplicated block split that block in two and lost the finding.

Measured on this repo's own test suite:

approach	clones removed	genuine duplication lost
mask literal tokens	58	~50
span rule (shipped)	5	0

A/B of the pre- and post-fix binaries over identical pristine source: 157 → 152 clones, 0 new. Four removals are overlapping-window artifacts; the fifth is a real []domain.Clone{…} fixture table. Self-scan under the repo's own config is unchanged (13 clones, 5.1%), so the CI gate is unaffected.

Language coverage

Go ([]T{…}, map[K]V{…}, []struct{…}{…}), Python, JavaScript, TypeScript, Ruby, Rust, PHP, Elixir, Swift, Dart, Lua.

Java, C, C++, C# and Kotlin are deliberately excluded — there [ is an index or an array type and { is also a block, so a literal cannot be told from code by tokens alone. Under-suppression is the intended failure mode: a missed false positive is noise, a swallowed real clone is a silent correctness loss.

Bugs found while building this

Fixed here rather than deferred:

• markPythonFunctions skipped every function after the first. It ended a body at the next blank-line-separated def, then resumed scanning after that keyword. Those bodies were never in scope, so their duplication was invisible. In a file of N defs, only the first was ever scanned.
• Dart bodies misread as literals. Bodies in languages without a func keyword come from a brace-depth heuristic that cannot tell a function body from a {…} map literal at the same depth. A real body follows ) or an arrow; a literal follows =, , or [.
• TokenizeFile panicked on a nil language while markFunctionBodies documents nil as "no syntax knowledge". Unreachable today (collectFiles drops unknown-language files first), but the contract was incoherent.

Tests

469 pass with -race; lint clean. Every new test was run against main and fails there:

• services/table_literal_test.go — the issue repro verbatim, plus gofmt-style rows
• services/overlap_test.go — the no-overlap invariant, and that disjoint repeats still report
• services/composite_literal_test.go — literal vs. logic, func type vs. func literal, barrier cases (empty, unterminated, start-of-stream, nil/opted-out language)
• services/data_literal_langs_test.go — per language: tables recognised, indexing (m[k]) not mistaken for a literal, callbacks inside tables still checked
• services/pyfuncboundary_test.go — 7 of 8 fail on main
• integration/table_literal_test.go — the built binary at real defaults; all 4 fail on main

Negative coverage matters as much as positive here: duplication in test files, in table callback columns, and in blocks containing inline literals is each pinned by a test, so "no clones" can't quietly mean "detection broke".

[github-actions]

✅ PR title follows the required format

Current title: fix: stop table-driven tests from being reported as clones