Blaise v0.9.0 (alpha) — Runtime port & bug squashing

Changelog — Blaise v0.9.0

Fixpoint: stage-2 == stage-3 at 166,479 lines of QBE IR (up from 135,311 in v0.8.0).
Test suite: 2,293 tests, 0 failures.

---

Runtime — C elimination

• Port blaise_arc_class.c to pure Pascal (blaise_arc.pas)
• Port blaise_exc.c to Pascal + x86_64 assembly (blaise_setjmp_x86_64.s
  remains as the only non-Pascal file)
• Port blaise_weak.c to pure Pascal
• Port blaise_float.c to pure Pascal Grisu1 float-to-string implementation
• Port blaise_str_fmt.c to pure Pascal _StringFormatN
• Consolidate blaise_io / blaise_process / blaise_sys / blaise_time
  into rtl.platform.posix Pascal unit
• Enable pipefail in the runtime Makefile so compiler errors fail the build

FFI / Codegen — interop correctness

• Records by value through SysV aggregate ABI: record-by-value parameters
  now use QBE :_ffi_<Name> aggregate types so fields are scattered into
  INTEGER/SSE registers per the platform ABI. Applied uniformly to all call
  sites (external and intra-Blaise), so callbacks are consistent. Win64 ABI
  comes from QBE for free. Previously, a bare l <addr> was passed and the
  callee read the pointer bits as struct contents.
• Single field alignment corrected to 4 bytes (was falling through to
  the 8-byte catch-all, doubling the size of Single-heavy records and
  breaking C struct interop)
• Double literal narrowed to Single on record field store — previously
  emitted stores d_<lit>, ... which QBE rejected as a type mismatch
• Double expression narrowed to Single before FFI call — a Double-
  typed actual passed to a Single formal put 8 bytes into the float slot;
  the C callee read the low mantissa half as IEEE-754 single (noise)
• Narrow FFI return values masked to declared width: Byte/Boolean
  returns are ANDed with 255, Word with 65535, SmallInt sign-extended
  via shl 16 / sar 16. C ABI leaves upper bits undefined for sub-int
  returns; callers observing the full word saw garbage.
• @A[i] on dynamic arrays — EmitAddrOfExpr now has a tyDynArray
  branch mirroring the open-array path. Previously raised "Unsupported
  L-value form for var argument".
• Overloaded Destroy in ARC field-cleanup — EmitFieldCleanupFn now
  calls the overload-mangled symbol (stored as DestroyResolvedQbeName on
  TRecordTypeDesc) rather than the bare <Class>_Destroy label, which was
  never emitted for overloaded forms, causing a linker undefined-symbol error.
• Fix storel for proc-pointer static-array slots; add Expr()() postfix call
• Use narrow load/store for implicit-Self Byte/Word/SmallInt fields
• Fix QBE types for Single float arithmetic and comparisons
• Emit correct data items for typed array consts inside function bodies
• Fix ARC retain/release on array element writes
• Allocate 16 bytes for procedure-of-object class fields
• Retain/release class refs stored through typed pointer
• Nil class/string field slots after .Free to prevent double-free
• Load class pointer before offsetting in var-param L-value
• Preserve ARC variables across try/finally re-raise

ARC — correctness

• finally blocks run on non-local exit (Exit, Break, Continue)
  via FFinallyStack — previously skipped entirely on non-local unwinds
• [Unretained] attribute for non-owning class references (back-pointers,
  AST Resolved*Type fields, symtab links) that must not participate in ARC
• Release +1-owned temporaries produced by function/property returns when
  used transiently or passed as value arguments
• Release +1-owned receiver temporaries after method calls
• Honour [Weak] on implicit-Self class field assignments
• Retain caught exception when binding to handler variable (except E: T do)
• Eliminate double-_ClassAddRef in constructor-with-args codegen
• Eliminate double-_ClassAddRef on function-return class assignments
• Wire owned-arg-temp release into all remaining call paths
• Remove 50+ redundant Free calls from AST destructors (field cleanup
  already handled by ARC)
• Remove redundant Free calls from symbol-table destructors

Language — new features

• Nested procedures with captured-variable closure (full access to
  enclosing scope locals)
• Custom attributes — [MyAttribute] / [MyAttribute(args)] on any
  declaration; compiler validates attribute class and resolves constructor
• [Unretained] non-owning reference attribute
• SmallInt and Word — 16-bit signed/unsigned integer types with correct
  narrow load/store and Byte/Word field packing at 1-byte / 2-byte stride
• UInt64 / QWord — 64-bit unsigned integer type
• packed record — suppresses per-field alignment padding
• sar arithmetic right-shift operator
• High / Low extended to all ordinal types; targeted float error message
• SizeOf(expression) — previously only SizeOf(TypeName) was accepted
• / is real division — distinct from div (integer division)
• for..in over dynamic arrays
• Octal / binary literals (0o777, 0b1010) and underscore separators
  in numeric literals (1_000_000, 0xFF_FF)
• Integer-type typecasts in const initialisers — Cardinal(-11),
  Byte(255), SmallInt(-1), etc., with correct bit-width truncation and
  sign extension
• Bit-op chains in const initialisers — FG_BLUE or FG_GREEN, 1 shl 8,
  $FF and 15, A xor B, including mixed literal+named chains and array
  elements; folded at semantic time
• ArcSin, ArcCos, Sinh, Cosh, Tanh math builtins with Single
  dispatch overloads
• Extend High / Low to dynamic array types
• Allow direct invocation of proc-typed class fields
• Allow implicit integer-to-float assignment; reject implicit float↔integer

Semantic — fixes

• Implicit Self member no longer shadows unit-level symbol
• Nested procs with the same name in different outer procs no longer ambiguous
• Clone method bodies per generic instance (was sharing mutable AST nodes)
• Allow two instances of the same generic class in the same scope
• Wire up parent class on generic instances
• Walk parent chain when resolving inherited properties
• Register var decls before analysing method bodies
• Use overload resolution for constructor calls
• Emit support data for generic classes declared in a unit
• Normalise all identifier references to declared casing
• Reject duplicate identifiers across const and var sections
• Report missing units even when no --unit-path is given
• Fix SizeOf(expression) to resolve the expression type

Threading

• TThread, TCriticalSection, and full POSIX thread bindings
• Self-referential thread lifetime under ARC (thread holds own refcount during
  Execute, preventing use-after-free on fire-and-forget patterns)
• [Threaded] E2E test suites run as parallel subprocesses

Stdlib / Performance

• GetText and GetCommaText rewritten to avoid O(N²) string concatenation
• TList<T>.IndexOf added
• TListEnumerator<T> + TList<T>.GetEnumerator — for..in over generic lists
• _ClassRelease elision for provably-nil slots (codegen perf)
• Remove all dead {$IFDEF FPC} compiler directives

Developer experience

• --debug leak reporter — shows refcount per live object at program exit
• blaise.cfg — automatic unit-path discovery from project directory
• Multiple --suite filters accepted by the test runner
• --emit-ir path now exits through normal cleanup instead of Halt(0)
• UTF-8 bytes in comments: regression tests added to lexer suite

Compiler internals / Build

• Mark Resolved*Type AST fields [Unretained] — prevents spurious retain
  cycles on type descriptors shared across the AST
• Remove all old FPC compiler directives no longer needed in Blaise
• --cache-dir flag and migrate_full.py removed (dead code)