core: shield string literals in positional comment stripper (#259)

[squid-protocol]

What

Prism._strip_positional_comments (the positional_anchored / positional_abap families — cobol, fortran, abap) split each line directly on *> / ! / " with no string-literal shielding, unlike every other stripping path in prism.py (_strip_single_line_comments, _strip_segment_comments, _strip_nested_comments all mask via LITERAL_MASK_PATTERN first).

A delimiter-shaped character inside a string literal was misread as a real inline-comment marker:

_strip_positional_comments('       DISPLAY "Rate *> 5%" TO CONSOLE.\n       MOVE X TO Y.', cobol_mode=True)
  # before:  code = '       DISPLAY "Rate \n       MOVE X TO Y.'   ← "TO CONSOLE." lost, dangling unterminated "
  # after:   code = '       DISPLAY "Rate *> 5%" TO CONSOLE.\n       MOVE X TO Y.'

Same failure for the ! branch (DISPLAY "Warning!") and the ABAP " branch (x = 'he said "hi"'). The tail of the literal + rest of the statement moved to the comment stream, deflating coding_loc and silently dropping real structural signals.

The prior closure ("resolved by #273") referenced the wrong PR — #273 only touched network_risk_sensor.py. #1745 fixed a different bug in the same function (line-count geometry). No regression test guarded this.

Fix

• Mask string/char literals per line (the same bounded discipline _strip_single_line_comments adopted for Contiguous 'dead zones' of real functions silently lost from extraction on production files, independent of mid-file language drift #1184) before the inline *> / " / ! search, restoring them into both the code and comment halves after the split.
• The column-1 / column-7 anchor check stays on the raw line so column positions are exact.
• ABAP masks single-quote / backtick only — its " is the comment delimiter, not a string quote (new ABAP_LITERAL_MASK_PATTERN).
• _mask_line_literals gains an optional pattern override (backward compatible).
• Out of scope: a literal continued onto a later line via a fixed-form column-7 - indicator — masking is deliberately line-bounded (Contiguous 'dead zones' of real functions silently lost from extraction on production files, independent of mid-file language drift #1184). Noted in the code comment.

Golden masters (re-blessed, both modes)

python tests/tools/update_golden_master.py run once per mode. 5 diffs, all on one file: cobol/che-che4z_lsp_project_fixtures/special_schema.cbl, whose DATA ... VALUE clauses hold XML-ish string literals containing ! (' <itemName>Not here!</itemName>'). With the literal no longer truncated at the !:

• Type/Safety Bypasses: 0 → 1 (a real signal previously hidden in the mis-stripped comment portion)
• Error & Exception Exposure: 62.98% → 65.52% on that file, + 3 aggregate rollups

No line-shift noise anywhere else across the corpus.

Verification

• ✅ tests/core_engine/ (572 passed) + new test_prism.py cases (delimiter-inside-literal for cobol *>/!, fortran !, abap ", plus real-inline-comment control cases)
• ✅ cobol / fortran / abap extraction gauntlets + _strict siblings (412 passed)
• ✅ tests/tools/crucible_check.py — both modes PASS after re-bless
• ✅ tri_comparison_chart.py --all --ci, tree_sitter_accuracy_audit.py --all --ci (pre-existing unrelated css corpus-drift failure reproduces on clean main with these changes stashed — not introduced here)
• ✅ ruff_audit.py --ci, mypy_audit.py --ci, dead_key_audit.py --ci, ruff format --check on prism.py

Closes #259

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab

[squid-protocol]

The failing tree-sitter-accuracy-audit / css check is pre-existing and unrelated to this PR — real_functions 24→25 drift from a stale css baseline (last regenerated in #2478; #2500's audit-reader changes weren't re-blessed). Verified it reproduces on main (4cc5001f) with this branch's changes stashed, and #2500 merged with the identical failure. This PR only touches _strip_positional_comments (cobol/fortran/abap), which cannot affect css.

Tracked separately in #2507 (also covers the @layer/@media/@supports-as-function FP the same audit surfaces).

All other checks green: crucible-audit both modes, full-suite-matrix ×40, tri-comparison, ruff/mypy/dead-key.