tools: refactor code-quality scorer and add header/macro/PARAM rules

[mohanchen]

Summary

Refactor tools/03_code_analysis/code_quality_score.py and extend it
with new AGENTS.md-backed rules. All changes are confined to the
scoring tool and AGENTS.md; no solver or build code is touched.

Changes

1. Split the scorer into focused modules (bfecaf3cb)

The 2349-line code_quality_score.py now delegates to five
single-responsibility modules under tools/03_code_analysis/, while
the entry command (python3 code_quality_score.py ...) is unchanged:

Module	Responsibility
cqs_config.py	weights, caps, thresholds, compiled regexes, prefix helper
cqs_text.py	comment/string stripping, comment-block extraction
cqs_rules.py	rule-detection (find_* / analyze_*) functions
cqs_analyze.py	per-file analysis orchestration, file discovery
cqs_report.py	Finding/FileReport dataclasses, text/JSON renderers

This commit also folds in two rule additions made just before the
split and broadens one existing rule:

• complex_argument_expression (AGENTS.md §14): penalises a ternary
  ?:, && or || used as a top-level call argument (weight 1, cap 5).
• ifdef_in_call_args: penalises #if/#ifdef blocks wrapping
  arguments inside a call's argument list (weight 1, cap 5).
• the using-namespace rule now matches any using namespace X;
  directive instead of only using namespace std;.

2. Header-specific rules (f005fb6c4)

• header_too_long: a .h file exceeding 120 lines costs 1 per extra
  10-line block, capped at 5 (stacks with the generic 500-line rule).
• header_function_body: a non-inline, non-template function body
  exceeding 20 lines in a .h file costs 1 per extra 10-line block,
  capped at 5. inline functions, template <...> functions, and
  bodies at or below 20 lines are exempt.

3. New AGENTS.md rules 16 and 17 (ebb2716ca)

• Rule 16 — new_macro_definition: a #define NAME ... in a
  .h/.cpp file costs 3 per definition (cap 10). Guard-style macros
  (name ending in _H/_HPP with no parameter list or replacement
  text) and #ifdef/#ifndef usages of existing macros are exempt.
• Rule 17 — param_member_init: a member variable initialized from
  PARAM in a .h/.hpp file costs 5 per declaration (cap 10).
  Function default arguments and constructor member-initializer lists
  are out of scope (already covered by default_parameter /
  global_dependency).

Verification

No ABACUS compilation or runtime test was run (tooling-only change);
all checks are Python-level:

• python3 -m py_compile passes on all touched modules after every
  commit.
• Split commit: smoke sample (ternary/&&/|| args + #ifdef-wrapped
  args) scores identically before and after the split (92; rule A hits
  L7-L11, rule B hits L22, no false positives); a full source_base/
  scan (284 files) produces the same module rollup, text/JSON and -o
  output as before the split.
• header_function_body: a smoke header with a 23-line plain body, a
  23-line inline body, a 23-line template body and a 6-line body
  charges only the plain long one; source_base/matrix.h shows no
  false positive.
• new_macro_definition: a smoke header with a guard macro, an
  object-like macro and a function-like macro charges only the latter
  two; source_md/msst.h (guard only) shows no finding.
• param_member_init: a smoke header with = PARAM..., {PARAM...},
  static const = PARAM..., a method default argument and a
  constructor init list charges only the three declaration-site
  initializations; source_base/matrix.h and source_md/md_base.h
  show no finding.

Notes

• The entry command and CLI flags are unchanged; existing CI/pre-commit
  invocations of code_quality_score.py keep working.
• Known limitation (documented in code): inline written on the line
  above a signature is not detected by header_function_body.

[AsTonyshment]

There are a few small edge cases in the new checks that need refinement.