feat(dragonbox): add LLAR formula

[MeteorsLiu]

Add an LLAR Formula for jk-jeon/dragonbox version 1.1.3.

The implementation includes:

• Install the upstream header through the verified Conan include-directory patch.
• Build the static library with the evidenced fPIC option and publish its link metadata.
• Add an independent C++ consumer test covering dragonbox::dragonbox_to_chars.
• Validate exact-version selection, the lower unsupported boundary, option rejection, and cache-hit consumer behavior.

This completes the closed Dragonbox translation issue with a source-backed Formula and reproducible consumer validation.

Closes #111

[fennoai]

Review: jk-jeon/dragonbox 1.1.3 formula

Solid, well-scoped formula that closely follows the sibling patterns (recp/cglm, json-c/json-c). Verified as correct:

• The include-directory patch is right: upstream 1.1.3 installs headers under include/dragonbox-1.1.3/...; the patch flattens that to include/dragonbox/..., matching the consumer's #include "dragonbox/dragonbox_to_chars.h" and the -Iinclude flag. The a/a b/b prefixes strip correctly under git apply -p2, and --unidiff-zero matches the zero-context hunk.
• dragonbox_to_chars is genuinely a static library, so fPIC legitimately changes installed object code — the option dimension and the defaults/filter pattern (mirroring cglm) are justified.
• The consumer test builds in _llar_consumer, separate from the _build scratch tree, and derives flags from installDir rather than build metadata — correct cache-hit-safe pattern.
• Error propagation uses idiomatic ! / lastErr!, and versions.json correctly declares no deps.
• Security review found nothing: permissions, path construction (all from trusted ctx.* + literals), and license handling are all sound.

The inline notes below are minor/consistency items. A couple (the patch idempotency and the misleading comment) are worth addressing before merge.

Minor, not inlined:

• DRAGONBOX_ENABLE_SUBPROJECT=false (line 55) already matches the upstream default (OFF); it's harmless as documented intent but is a non-source-required flag per the formula-semantics guidance.
• The metadata flags (lines 67–72) and the test flags (lines 84–91) are hand-maintained duplicates and must be kept in sync manually. Acceptable since upstream ships only dragonboxConfig.cmake (no .pc), matching the cglm fallback pattern.

[fennoai]

[P3] Comment conflates two unrelated build choices (fPIC vs subproject)

This comment sits on the defaults block (which only defines fPIC) but ties the fPIC choice to DRAGONBOX_ENABLE_SUBPROJECT via "so expose ... while keeping the source's subproject disabled." They're independent: fPIC -> CMAKE_POSITION_INDEPENDENT_CODE (line 53), whereas DRAGONBOX_ENABLE_SUBPROJECT (line 55) only controls building the upstream subproject/benchmark/test and is never set by the Conan recipe. Suggest stopping the comment at the fPIC rationale and moving any subproject note next to line 55.

[fennoai]

[P3] versions.json uses 2-space indent; all siblings use tabs

jk-jeon/dragonbox/versions.json is 2-space indented, whereas json-c/json-c/versions.json, recp/cglm/versions.json, and madler/zlib/versions.json all use tabs. Match the sibling convention (tabs) for consistency and to avoid noise in any format check.

[fennoai]

Review (inline notes, part 2)

Additional inline findings from the dragonbox 1.1.3 review; the summary is in the preceding review.

[fennoai]

[P2] Patch is non-idempotent; scratch file dropped into CMake source root

onBuild writes _llar_dragonbox.patch into ctx.SourceDir (the CMake source dir passed to cmake.new on line 52) and runs git apply --unidiff-zero -p2 on every build. Two concerns:

1. Not idempotent: git apply has no --forward/reverse guard here, so if the source tree is ever reused without a clean re-extraction (e.g. an interrupted prior build that already applied the patch), the second apply fails on the already-patched tree. Consider git apply --reverse --check (skip if already applied) or git apply --check first.
2. Scratch file at source root: writing the patch into ctx.SourceDir can perturb CMake configure on generators that file(GLOB ... CONFIGURE_DEPENDS). Siblings keep scratch under subdirectories and don't drop loose files at the source root — writing the patch into _build (or a temp path) avoids this.

[fennoai]

[P3] Header-flatten patch is a non-obvious deviation with no comment

The patch flattens upstream's versioned header install path (include/dragonbox-1.1.3/...) to include/dragonbox/... so the plain -Iinclude flag and the consumer's #include "dragonbox/dragonbox_to_chars.h" resolve. This is deliberate but unexplained at the application site. Sibling formulas document their install-layout assumptions inline (e.g. Cglm_llar.gox, Jsonc_llar.gox); a one-line comment here would keep a future maintainer from mistaking the patch for cruft.

[fennoai]

[P3] Hardcoded license filenames hard-fail if upstream renames them

This is the only formula in the repo doing manual license copying, and the two names are hardcoded. If either is renamed across any future fromVer range extension, os.readFile(...)! aborts the build with an opaque error. Names are correct for 1.1.3, so this is fine as-is — but confirm this licenses/ layout is a required store convention (siblings copy none); if nothing consumes it, consider dropping the block, and if it stays, note the tight coupling to exact upstream filename spelling.