fix(tests): pair same-named function occurrences by line instead of collapsing them

[squid-protocol]

Summary

• tree_sitter_accuracy_audit.py's per-file comparison keyed same-named function occurrences by
  name alone in a plain dict. A file with multiple same-named functions (property getter/setter
  pairs, __init__/__call__ across different classes, repeated methods) silently kept only the
  last tree-sitter occurrence and the first GitGalaxy occurrence -- comparing unrelated
  functions. This wasn't just an args_exact_match problem: real_functions/found_functions
  themselves were undercounted the same way whenever names repeated within a file.
• Persists function_data.start_line (gitgalaxy/recorders/record_keeper.py -- detector.py
  already computed this internally per function, it just never made it into the DB), and uses it
  in the audit tool to pair same-named real/found occurrences by an order-preserving
  minimum-line-distance alignment (_align_occurrences_by_line) instead of collapsing them onto
  one dict slot.
• Also excludes detector.py's own synthetic Anonymous_Block/__global_context__ placeholder
  names (used for top-level control-flow blocks in script-style files with no enclosing function)
  from the comparison -- no real language grammar can produce either literal name, so counting
  them as "extra" isn't measuring precision. Left uncorrected this manufactured a wildly misleading
  18.0%/10.7% func precision reading for matlab/shell once same-name collapsing stopped hiding it.
• Regenerated all 31 language baselines and language_standards.py's summary table to match the
  corrected methodology (this file is shared across every NODE_MAPS language, so the fix's effect
  isn't python-scoped even though tree_sitter_accuracy_audit.py: duplicate function names within a file collapse in the per-file comparison #1526 was found via python).
• Documented two now-visible categories of "GitGalaxy beats tree-sitter" (Cython cdef class
  scope-loss, MATLAB/shell synthetic-block category mismatch) in
  docs/why_gitgalaxy_beats_ast_here.md, and added a standing CLAUDE.md instruction to log any
  future case there.

Two anomalies filed separately, not fixed here

Regenerating the baselines surfaced two patterns that look like real detector.py extraction
bugs rather than audit-tool artifacts, but are out of this PR's scope (methodology fix, not an
engine-correctness hunt):

• Fortran: detector.py's start_line is wrong for some subroutines, ~370 lines off in module_initialize_real.F #1531 -- Fortran: start_line is ~300+ lines wrong for a couple of WRF subroutines
  (wrf_message, const_module_initialize), with args also wrong for the same functions --
  reads like func_start's span anchored somewhere other than the real declaration.
• Haskell: detector.py's start_line looks shifted by roughly one prior function when short definitions are stacked #1532 -- Haskell: start_line looks consistently shifted by roughly one prior function
  (~10-13 lines) when several short name :: Type / name args = body definitions are stacked
  back-to-back.

Test plan

• python tests/tools/audit_check.py -- ruff/mypy/dead-key/ast-accuracy all clean
• python tests/tools/tree_sitter_accuracy_audit.py --all --ci -- all 31 languages pass
• python tests/tools/tree_sitter_accuracy_audit.py --summary-table -- no diff (already
  regenerated and committed)
• python -m pytest tests/core_engine/ tests/tools_recorders/ tests/extraction/ tests/cobol_mainframe/ -q -- all pass (2 unrelated tests/security_auditing/ runs timed
  out locally at 100-115s with no failures observed before timeout; unrelated to this PR's
  files, not investigated further)
• Verified start_line lands correctly end-to-end with a real galaxyscope --db-only scan
  against a hand-built sample file before wiring it into the audit tool
• Not a parsing-logic change (detector.py/prism.py/language_standards.py's rules are
  untouched; only the summary table, itself derived from baselines, changed) -- no differential
  scan / golden master regen needed

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab