Cut the arg-parser generator over to the embedded runtime

[Smaug123]

Stacked on #566. The cutover PR of the arg-parser rewrite.

What changes

Generated parsers no longer contain a hand-emitted scanning state machine (~670 lines of quoted-AST emission deleted). Instead:

• The property-tested runtime ships inside the generated file. ArgParserRuntime.fs is carried as an embedded resource, parsed at generation time, and spliced into generated output as a private module (one per namespace per input file, named ArgParserRuntime_<first tagged type> — a name that cannot collide across generated files, since the type itself would collide first; the prefix is reserved in the same sense the generated per-type module names always have been). Generated code stays dependency-free, and the exact text shipping in every consumer is the code the FsCheck suites exercise in this repository.
• Per type, the generator now emits only data + typed glue: an ErasedSchema value describing the parser's shape, converter/storage callbacks over the typed slots, and record assembly. The runtime's runParse (new in this PR, unit-tested over fake callbacks) drives scanning, routing, duplicate/missing detection, structural message rendering, and defaults, keeping diagnostics in argv order. Selection is shape-first by construction: a conversion failure can never change how a token was routed.

What doesn't change

Every pre-existing arg-parser test passes unchanged — help formats, negation semantics, keyed positionals (under every long form), env-var defaults, ParseExact, extension-methods mode, and the deliberately-pinned eccentricities (greedy value consumption etc.).

Deliberate semantic changes (= the triaged bug fixes; TestArgParserKnownBugs flipped accordingly)

• A malformed --key value no longer aborts the scan (errors carry (at arg ...) context).
• A duplicated flag no longer swallows the following token. A duplicate of a slot whose first occurrence failed conversion is not an error at all: the later occurrence fills the slot, exactly as before the rewrite.
• List, positional, post-separator, and env-var conversions feed the aggregated Errors during parse! channel instead of escaping as raw FormatException.
• Environment lookups no longer run once the parse has failed.
• --help matches case-insensitively, like every other argument.
• A non-positional bool list field now generates compiling code (new NonPositionalBoolList regression type; previously uncompilable).
• Duplicate messages distinguish valued occurrences (Argument '--foo' was supplied multiple times: 3 and 4) from bare flags (Flag '--baz' was supplied multiple times).

Explicitly not addressed here (unchanged behaviour, follow-ups)

• Generation-time collision validation is still case-sensitive while runtime matching is case-insensitive.
• Nested-record default functions are still invoked on the root type.
• The embed round-trip constraint (no char literals / _.member shorthand in the runtime file) is enforced end-to-end by CI's regenerate-and-build check.

🤖 Generated with Claude Code

[Smaug123]

Review round 2 addressed:

• False duplicate diagnostics after conversion failures: duplicates are now detected in runParse's fold from the first successful store, matching the historical parser exactly — --foo=bad --foo=1 reports only the conversion error and stores 1; --foo=1 --foo=2 reports "1 and 2". Regression test added.
• Runtime module name collisions with user declarations: consciously not chased further. An untyped-AST generator cannot see declarations in other input files, so no name is provably collision-free — the same already holds for the generated per-type modules themselves (Foo/FooArgParse). ArgParserRuntime_<Type> is hereby a reserved prefix, like those. What is guaranteed: two generated files can no longer collide with each other (module named after a tagged type rather than the file).

🤖 Generated with Claude Code