feat(tri-comparison): superscript markers for args-granularity mismatches

[squid-protocol]

Summary

• "args" doesn't mean the same unit across every language, even when a real comparison tool exists for other metrics -- COBOL's Args Found panel was silently rendering nothing at all (0/126 real args comparisons ever possible, since ctags has no per-paragraph signature for COBOL either), which read as "no data" rather than "this language measures args differently by design."
• Adds ARGS_GRANULARITY, a small per-language classification with its own superscript marker + legend line (same convention */** already use):
  • program_level (†): one whole-program signature, not per-callable. cobol: PROCEDURE DIVISION USING/RETURNING is a program-level header; paragraphs don't take individual arguments at all.
  • none (‡): no parameter-list concept exists, ever. dockerfile/yaml/jcl: func_start matches a document-structural marker (RUN/CMD/EXEC/script:) as a pseudo-callable, same reason non-function-shaped languages get a func_start rule at all.
  • proxy (§): a real, non-trivial measurement via something other than a parenthesized parameter list. assembly/agc_assembly: counts calling-convention register mentions as an arg-count heuristic, same spirit as the existing bash/Perl $1/$2/$3 precedent.
  • default (no marker): ordinary per-callable parameter-list count.
• The classification and the "don't guess -- rule out a real recall bug first" methodology behind it live in .claude/skills/tri-comparison-ledger-sweep/SKILL.md's "Args granularity" section (docs(skill): document args-granularity classification in tri-comparison-ledger-sweep #1924); this PR is the chart-rendering implementation of that design.

Verification

• Regenerated the full 45-language chart, diffed by label TEXT CONTENT (not raw SVG, since bumping top_margin by 14px for the new legend line shifts every row's y-coordinate) -- confirmed exactly 4 new bar labels appeared (cobol 14†, dockerfile 0‡, assembly 77§, agc_assembly 469§) and exactly 1 new legend line, nothing else in the chart changed.
• Confirmed the ranked-panel summary tally is untouched (Args Found stays unranked per the module's own PANELS docstring).
• ruff_audit.py --ci / mypy_audit.py --ci: clean, no new findings (pre-existing, unrelated formatting drift in this file confirmed present even on a clean checkout with my changes stashed out, left untouched per repo convention for pre-existing debt).

Test plan

• Test-rendered a small language subset first (cobol/dockerfile/assembly/agc_assembly/python) to confirm markers render and defaults are unaffected, before the full --all --write run
• Full chart diff verified by label content, not raw coordinates
• Confirmed unrelated stray ledger.json churn (non-deterministic anonymous-name IDs in capped examples, from running the full reconciliation) was reverted out of this PR -- kept focused on the chart-rendering change only

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab