fix: make task wait dominance path-complete

[sanohiro]

Summary

• replace traversal-order task wait state with group/generation/epoch-scoped dominance proofs
• preserve successful Wait and Move Task provenance through transparent control flow, joins, loops, and nested groups
• reject TaskGet until the originating generation is completed, while keeping the TaskGroup ABI unchanged

Verification

• scripts/test-pr.sh (LLVM 22 via /opt/homebrew/opt/llvm, isolated CARGO_TARGET_DIR)
• cargo clippy -p align_sema --all-targets --locked -- -D warnings
• cargo test -p align_sema --lib (188 passed)
• focused task-group owner tests, including repeated Spawn, stable join tokens, and branch/loop dominance cases

The implementation is intentionally one vertical because splitting proof state, transport, joins,
or TaskGet diagnostics would expose an intermediate compiler that could authorize an uninitialized
task slot or reject a valid outer proof. The split rationale is recorded in HANDOFF.md.

Post-open host-native review: CLEAN. A fresh independent adversarial review also found no actionable
findings after the stable syntactic-site join-token fix. The native host transcript is preserved
under .git/amw-host-final-native.log.