fix(core-engine): scheme globals decided by enclosing form, not indentation (#2674)

[squid-protocol]

Closes #2674.

What

Scheme's globals regex matches every (define name value), but in Scheme indentation does not track scope: (define y 5) inside (define (f x) ...) is a local binding and (define y 5) inside a file-wrapping (let () ...) is a real global, at identical indentation. On the language-crucible Chez/Racket sources the bare regex is 42.7% precise, and the #2651 column-0 anchor would have deleted 32 real globals along with 67 false positives (two of five real files have no column-0 defines at all).

This adds a small, registry-declared scope filter hook to coding_analysis and one filter, lisp_body_position:

• gitgalaxy/standards/language_standards/languages/scheme.py declares "_scope_filters": {"globals": "lisp_body_position"} (a private key, skipped by every rules iterator the same way _args_* / _dependency_capture already are).
• detector.py runs the filter after the regex over the same segment. It only ever removes matches, so counts, spatial_map and threat_locations stay consistent; an unknown filter name is ignored with a diagnostic, so a registry typo can never zero a metric. The structural pass is cached per segment.
• _lisp_module_level_define_offsets is a linear tokenizer + form-head stack (strings, #\( char literals, nested #| |#, [ ] all handled). Each define climbs to its nearest non-transparent frame: a module frame (library/module/define-library, or the outermost bindings-less let-family wrapper, either only when no body frame sits above it) keeps it; a body frame (lambda / let-family with bindings / when / cond / … / any define-*) drops it; begin, let-syntax (R6RS 11.18 splicing), if, cond clauses and quoted data are transparent.
• how_to_add_a_language.md gets engine rule 17: when the discriminator is the enclosing form, declare a scope filter rather than faking it with a column anchor.

Second defect found on the way (Prism, same PR)

#\; is a Scheme char literal ((write-char #\; p), cpnanopass.ss:7690/7805, io.ss:6393). _strip_nested_comments's ; line-comment branch claimed it and swallowed the rest of the line including its closing parens, so every paren-balanced scan downstream ran one level deep for the rest of the file — the first cut of this filter read cpnanopass.ss as 0 globals because of it. Fixed by claiming the char literal atomically before the comment branch (gated to recursive_block_lisp, same mask/unmask path as strings). Pre-existing; it also affected the Mode-B function slicer.

Measurement

Golden master (Prism code streams, i.e. what coding_analysis actually sees); full-precision and zero-dependency move identically:

file	globals before	after	the dropped defines sit in
cpnanopass.ss	26	1	define-pass bodies (12), define-who/lambda bodies (3), meta-cond/with-output-language clauses (8), a local (module …) under a block (let () …) (2)
io.ss	10	6	(define who '…) in procedure bodies and a nested (let () …) block; the 6 kept are the wrapper-level buffer-size constants and open-files
schemify.rkt	6	0	cond/match clause bodies and procedure bodies
thread.rkt	3	1	procedure bodies; the kept one is a top-level callback

Plus one cell from the Prism fix alone: io.ss Closures and Anonymous Functions 454 → 455 (the (lambda () #\;) at io.ss:6393 the stripper used to eat). Attributed by re-scanning with each half alone: Prism-only moves exactly that one cell; the four globals cells belong to the filter.

Against the issue's raw-source paren-depth oracle (53 kept / 71 dropped) the walk deliberately differs in three shapes the oracle's "depth == 1, or any module" proxy could not express:

• kept — defines under a (begin …) or a let-syntax body inside the wrapper: both splice into the enclosing context (R7RS 5.6.1, R6RS 11.18), so they have the same status as the wrapper's direct defines;
• dropped — defines inside a (module …) that itself sits inside a nanopass define-pass (a procedure): pass-local state;
• dropped — defines inside a (module …) nested under a block-scope (let () …): block-local, the same rule the oracle itself applied to syntax.ss's (define who …) runs.

(The issue's raw-source counts also include defines whose "value" was a trailing ; comment, which the code stream never shows the regex — so the golden-master figures above, not the 53/71, are the right bar.)

Verification

• tests/extraction/languages/test_scheme_strict.py: 13 new end-to-end tests through StructuralExtractor.splice (issue example, 12 body forms, file wrapper, nested block let, begin/let-syntax splicing, library/module/define-library, local module, string/char-literal desync guard, square brackets, count/spatial/location consistency, unknown-filter fallback, linear-time check on 7 pathological payloads incl. 200k parens and an unterminated string).
• tests/core_engine/test_prism.py: #\; / #\" / #\space survive the stripper verbatim, the real comment after them is still stripped, paren balance preserved.
• Full suite green; ruff format; ruff / mypy / dead-key audits report nothing new.
• scope_check.py --expect scheme: all golden-master movement confined to scheme in both modes.
• tree_sitter_accuracy_audit.py --ci --all: 30/30 OK. tri_comparison_chart.py --ci --all: all OK.
• keyword-rosetta verify_language.py scheme: PASS (72 assertions) before and after — the corpus plants at top level and cannot see this defect, so no re-bless is owed.
• Golden master: not yet re-blessed — the five cells above need crucible_check.py --update --yes, which I leave to the user per repo convention.

Cross-repo

None. keyword-rosetta's scheme gate is unchanged against this build; no corpus PR and no ledger entry (the issue's fallback of ledgering scheme globals as a known over-count is superseded by the fix).

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab