Restore drop-in compatibility: make 'Release' a contextual keyword

[CoreyRDean]

Summary (non-technical)

Classic Blitz3D programs freely used release as an ordinary variable name. BlitzForge added a Release keyword (the GC release operator) and reserved it globally, so any legacy program with a release variable failed to compile with a baffling "Expecting identifier" — a direct hit to the project's #1 promise: compile their old game on day one, no source changes. This makes Release a contextual keyword: it's the GC operator only where it actually is one (Release.Type <obj>), and an ordinary identifier everywhere else. This is Phase 2 of the contextual-keyword work that #77 began with Test.

Technical summary

The Release.Type operand operator always has a type identifier immediately after the keyword (with an optional .). So Release is the keyword iff the next token is . or an identifier; otherwise it's a plain identifier (release=, release\f, release(, release[, If release=1). Applied at the same three parser.cpp sites #77 used for Test:

• parseIdent() now accepts RELEASE → release works as a parameter / Local / Global / Field name and type-instance var.
• Statement dispatch routes release = … / release(…) / release\field through the IDENT statement path; Release.Type x keeps stmtTok==RELEASE.
• Expression primary handles Release[.]Type operand when a ./identifier follows, else falls through to the IDENT case.

No tokenizer change; ReleaseNode semantics and the Release.Type grammar are untouched.

Scope is Release only. The other deferred token, Object, is a base Blitz3D keyword (Object.Type(handle)), so object-as-identifier wouldn't have compiled in stock Blitz3D either — a different problem class, left deferred with an explanatory allowlist note.

No breaking changes.

Acceptance criteria & results

• release compiles as Local, assignment target, If-condition operand, function parameter, and field-typed var — pinned in tests/LegacyKeywordIdentifierTest.bb; blitzcc -t passes (5 new blocks).
• The Release.Type obj GC operator still parses and works — positive test asserts RefCount goes 1 → 0 after release.
• samples/Blitz 2D Samples/spindisc.bb compiles; its allowlist line is removed.
• Corpus sweep: 352 files, 326 compiled (was 325), 26 known-fail (was 27), 0 new-fail.
• Full test.bat green ("Tests passed").

Trade-offs / deferred

• Object contextual-identifier handling (base-keyword problem class — needs separate investigation).
• Recast/Reference share the same expression-position grammar shape; no corpus collision today, but the same pattern now applies if ever needed.
• Strongest next-iteration candidate: modern-features docs as a CI-gated executable contract + BBList documentation.

🤖 Generated with Claude Code