codedna v1.0.3

Accuracy release. Every change below corrects a measurement or a claim that was wrong, so profiles generated with 1.0.2 may report different numbers after upgrading. That is the point.

Fixed

• Fix comment density and code-line counts for Python files that open a multi-line string mid-line (SQL blobs, templates, prompt constants). The closing triple quote was read as the start of a docstring, so the remainder of the file was counted as comment. On codedna's own repository this reported 13.4% comment density against a true 0.9%.
• Count multi-line JSDoc blocks as documentation. Only single-line /** ... */ was recognized, understating doc-comment coverage on any codebase that uses conventional JSDoc.
• Stop counting a plain // aside as a doc comment outside Go. Go keeps //, since that is its documentation convention.
• Measure function length in tab-indented Python. Every function in a tab-indented file was reported as one line long.
• Read files with a UTF-8 byte order mark so the first line is not hidden from every pattern.
• Skip minified and generated files, which dominated naming and identifier-length histograms and could invert the reported dominant convention. Skipped files are disclosed in the output.
• Require a word boundary for boolean prefixes, so issue and island no longer count as is-prefixed.
• Do not count blank lines inside a block comment toward a metric defined as a percentage of non-blank lines.
• Match the <!-- codedna:start --> and <!-- codedna:end --> markers as a pair when wiring. A stray closing marker earlier in the file made every run append another block.
• Preserve indentation after the wired block, and leave instruction files byte-identical when re-wiring. The second run on a fresh repository previously produced a spurious diff.

Changed

• Note in the stats output when UPPER marks a single-word all-caps constant compatible with SCREAMING_SNAKE, mirroring the existing lower note.
• Correct CONTRIBUTING.md, which presented the wiring helper as a read-only sanity check. It writes files, and the instructions now say so and point at a throwaway copy.
• Correct docs/AGENT_SUPPORT.md, which said AGENTS.md is created only when no other instruction file exists. It is created or updated either way.

Added

• Add regression tests for every fix above, doubling the suite from 15 tests to 33.
• Add a test asserting the version in skill/SKILL.md matches every other version reference in the skill and the agent support docs, including the release tarball URL.
• Accept start_marker and end_marker arguments in replace_block. Nothing passes them yet; they exist so a sibling skill can reuse the helper without its markers defaulting to codedna's.