wasm: nested constructor patterns in match arm args (#68)

[hyperpolymath]

Closes #68.

What changed

compile_match in ephapax-wasm no longer uses br_table. Each arm is a Block($arm_i_fail) containing an optional tag check (for top-level Constructor patterns) plus a recursive compile_sub_pattern walk over the destructure tree. Refutable sub-patterns (Literal, nested Constructor) emit br_if $arm_i_fail on mismatch, falling through to the next arm.

compute payload + flat_tag (only for n>=2 data scrutinees)
block $done : i32
  block $arm_0_fail
    (tag check if ctor; br_if $arm_0_fail on mismatch)
    (destructure args via compile_sub_pattern;
     refutable sub-pat → br $arm_0_fail)
    (arm body)
    br $done
  end
  ...
  unreachable    ; typechecker proves dead for well-typed
end

Why drop br_table

br_table commits one outer-tag-value to one arm. When two arms target the same outer ctor with different inner patterns (Some(0) then Some(_)), the second arm is unreachable in the br_table form. With Maranget exhaustiveness from #66, the typechecker now accepts Some(Some(_)) + Some(None) + None as well-typed, but the br_table form silently miscompiles Some(Some(42)) for that layout. Linear dispatch costs O(N) tag checks instead of O(1) — negligible for typical N≤4 and required for correctness.

Sub-pattern coverage

compile_sub_pattern handles every Pattern variant:

Variant	Behaviour
Wildcard	no-op
Var	local.set to fresh local
Unit	no-op (single-inhabitant)
Literal(Unit)	no-op
Literal(Bool) / Literal(I32)	i32.ne + br_if fail_depth
Pair	load offset 0 / offset 4; recurse
Tuple	right-nested pair walk; recurse
Constructor	emit_tag_walk + tag check + recurse

destructure_payload walks the right-nested pair chain (offset 0 / offset 4) and recurses compile_sub_pattern for each constructor arg, sharing the same fail_depth.

Out of scope

• I64/F32/F64/String literal sub-patterns — currently route to immediate br_if 1 (fail). The typechecker accepts them but codegen needs i64.ne / fNN.ne / string compare. Follow-up issue (no number yet — happy to file if you want).
• br_table fast path for the fully-irrefutable case. Linear is correct; a fast path can be reintroduced once perf matters.

Test plan

• cargo test -p ephapax-wasm --lib → 73 pass (was 70), +3 new:
  • compile_module_match_nested_some_some — Option(Option(I32)) covered by None/Some(None)/Some(Some(v))
  • compile_module_match_nested_with_literal — refutable literal: Some(0) | Some(v) | None
  • compile_module_match_nested_tuple_ctor — data B = Box(Option(I32)) matched by Box(Some(v)) | Box(None)
• Each new test validates emitted bytes via wasmparser.
• No regression in the 4 prior compile_module_match_* tests from feat(typing,interp,linear,wasm): core ExprKind::Match — typing + interp + N-arm linearity + br_table codegen #65.
• Workspace lib tests: full sweep passes after AppLocker cache clears.
• CI green on all checks.

🤖 Generated with Claude Code