Introduces a single parser-agnostic settings struct and a layered
JSON pipeline that replaces the bespoke juce::ArgumentList parser.

Pipeline: each input layer (--config file, environment, CLI) becomes a
sparse nlohmann::json object; layers are merged with merge_patch in
precedence order (defaults < config < env < CLI, CLI wins) and
deserialised in one step into PluginvalSettings, which converts to the
existing PluginTests::Options at the JUCE boundary.

- PluginvalSettings.h: unified std-typed struct + toPluginTestOptions()
  + fromPluginTestOptions(); NLOHMANN_DEFINE_TYPE..._WITH_DEFAULT so
  missing keys fall back to defaults; RealtimeCheck enum<->string mapping.
- SettingsSerializer.{h,cpp}: JSON load/save + the comma-list, hex-seed
  and disabled-tests(file-or-list) coercions shared by the CLI/env layers.
- SettingsParser.{h,cpp}: preprocess (deprecation rewrite, macOS flag
  strip, implicit --validate), magic_args parse of the flat options into
  the sparse CLI layer, env layer, merge, and the child-process handoff.
- Child process round-trip now uses an authoritative base64 JSON arg
  (--config-base64) instead of per-flag re-serialisation, avoiding
  quote/tokenisation hazards. Round-trip asserted in tests.
- CommandLine.{h,cpp}: thin adapter; old manual parsers/env-merge/help
  text deleted. magic_args drives --help (auto usage + env-var trailer).
- New --config <file.json> option; precedence covered by tests.
- CommandLineTests.cpp rewritten against the new pipeline (defaults,
  parser, hex seed, comma lists, rtcheck, path resolution, implicit
  validate, env precedence, config precedence, round-trip).
- CMake: macOS deployment target 10.11 -> 13.3 (std::format in magic_args
  requires libc++ floating-point to_chars, available on 13.3+); add new
  sources.

All --run-tests pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- docs/Command line options.md regenerated from magic_args --help output
- CLAUDE.md: document the JSON-merge settings pipeline, new source files,
  magic_args + nlohmann/json deps, C++23, and the macOS 13.3 deployment target

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Addresses review feedback that the magic_args version duplicated the
settings into a parallel PluginvalCliArgs struct plus manual CLI->JSON
and ENV->JSON bridges (5-6 edit sites per new option), and forced
cxx_std_23 + a macOS 13.3 deployment target (std::format).

CLI11 binds each option directly to a PluginvalSettings member, with the
environment variable on the same line via ->envname(). Adding an option
is now three edits: a struct member, a nlohmann macro entry, and one
add_option(...) line. No parallel struct, no cliArgsToJson, no envLayer.

- SettingsParser: rewritten on CLI11; comma lists via ->delimiter(','),
  enum via CheckedTransformer, hex/int seed via a small callback. --config
  seeds the struct before parse so precedence stays defaults<config<env<CLI.
  --help/--version handled by CLI11 (auto usage + footer).
- SettingsSerializer: dropped the comma/JSON-array coercions; kept the hex
  seed + disabled-tests(file-or-list) helpers and JSON load/save.
- Child-process base64 JSON handoff (--config-base64) unchanged.
- CommandLineTests: env tests now use setenv/unsetenv (CLI11 reads the real
  environment); all other cases unchanged. All --run-tests pass.
- CMake: magic_args -> CLI11 2.6.2; revert cxx_std_23 -> cxx_std_20 and
  CMAKE_OSX_DEPLOYMENT_TARGET 13.3 -> 10.11.
- Docs/CLAUDE.md updated.

Verified: full build (validator ON, C++20, macOS 10.11), --run-tests pass,
real --validate SUCCESS, --config precedence covered by tests.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Three regressions surfaced by CI (develop is green):

- Linux build: std::int64_t (== long on LP64) vs juce::int64 (== long long)
  made expectEquals template deduction fail. Cast the settings int64 fields
  to juce::int64 in CommandLineTests. (macOS compiled because there the two
  types are identical.)
- "Pluginval as Dependency" Verify: verify_pluginval runs
  `pluginval --help | grep "JUCE v<version>"`. CLI11's auto help didn't print
  the JUCE version; add it to the help footer.
- Windows --run-tests exit 127: runUnitTests() called the [[noreturn]]
  juce::ConsoleApplication::fail(), which throws; outside a ConsoleApplication
  command handler that throw was uncaught -> std::terminate. Set the
  application return value directly instead.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- Linux build: include <CLI/CLI.hpp> before the JUCE headers. JUCE pulls in
  X11 on Linux (JUCE_GUI_BASICS_INCLUDE_XHEADERS), whose `#define Success 0`
  clashed with CLI11's CLI::ExitCodes::Success enumerator. macOS/Windows
  don't use X11, so only Linux was affected.
- runUnitTests now prints each failed test (name + messages) to stdout.
  juce::UnitTestRunner logs via juce::Logger, which on a GUI app (Windows)
  doesn't reach the console, so CI couldn't show which tests failed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

"Command line parser" asserted opts.dataFile/outputDir.getFullPathName()
against Unix-style literals; on Windows juce::File normalises those to the
current drive (e.g. "D:\path\to\file"), failing the comparison. Assert the
raw parsed settings strings instead, which are identical on all platforms.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Reworks the layering so a --config file (deliberate, per-invocation) ranks
above ambient environment variables, and makes --config repeatable.

- Environment layer: env-var names are now derived from the registered CLI11
  options (--strictness-level -> STRICTNESS_LEVEL) instead of CLI11's
  ->envname(), so there is no hand-maintained env table and env support is
  automatic for every option. A synthetic "--name=value" argv is built from
  the environment and parsed by CLI11, reusing all its coercion (comma lists,
  enum transformer, hex-seed callback). Flags accept --flag=1/0.
- --config: repeatable; each JSON file is merge_patch-ed in command-line order
  (later files win per key). Applied between the env and CLI layers, so it
  beats env and loses to explicit CLI options. (Inline JSON dropped for now to
  avoid command-line re-tokenisation hazards; file paths only.)
- The env layer reads through an injectable EnvProvider again, so the unit
  tests no longer mutate the real process environment (also removes the
  Windows _putenv_s fragility).
- Tests updated for the new precedence + a repeatable-config test.
- Help footer, docs and CLAUDE.md updated.

Verified end-to-end: env(6) -> 6; env(6)+config(2) -> 2; +CLI(9) -> 9.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- CHANGELIST.md: 2.0.0 entry summarising the CLI11/JSON parser rewrite,
  --config, the new precedence, and the breaking changes (VERSION not bumped).
- --config-base64 (internal parent->child handoff) now errors if combined with
  any option other than --validate, instead of silently ignoring it. Test added.
- SUBCOMMANDS_HANDOFF.md: handoff doc for the deferred subcommand restructure
  (validate/run-tests/strictness-help with deprecated flat-flag aliases).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- --validate-in-process is not a CLI option (launchInProcess is GUI/internal
  only); the new strict parser would now error on it. Removed from the options
  list and corrected the process-model notes (GUI uses child processes; the CLI
  --validate runs in-process with signal-handler crash reporting).
- Note the staged 2.0.0 CHANGELIST entry vs the un-bumped VERSION.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Source/tests/EditorTests.cpp and Source/tests/ExtremeTests.cpp were never
added to SourceFiles, so their self-registering PluginTest instances were
never compiled in and the tests never ran. Add them to the build.

Re-enables: "Editor stress" (L6), "Editor DPI Awareness" (L3, Windows only),
"Allocations during process" (L9), "Process called with a larger than prepared
block size". Verified they compile and pass a real strictness-10 validation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Qualify juce::ScopedDPIAwarenessDisabler. The EditorDPITest block is guarded
by #if JUCE_WINDOWS && JUCE_WIN_PER_MONITOR_DPI_AWARE, so this latent bug only
surfaced now that EditorTests.cpp is compiled (and only on Windows).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>