[tvolk131]

Shared aliases such as T0 = u8; Tn = (Tn-1, Tn-1) are stored as a compact graph, but structural lowering previously expanded them as a tree. This made type conversion, cast analysis, and compilation grow exponentially with alias depth.

Memoize completed subtype conversions by node identity within each conversion. Keep traversal iterative, and lower enum payloads through the same cache. Three regression tests check retained sharing, nested enum payload shapes, and 8,192 levels of nesting.

Validation

• Before/after encoded program hashes match for p2pkh, HTLC, and kitchen-sink fixtures, each with debug symbols enabled and disabled.
• Local release probe on a 466-byte, 18-level shared-alias program: analysis approximately 167 ms → 54 µs; compilation 336 ms → 90 µs. Best of three with allocation counting on an Apple M4 Pro; setup and output destruction excluded. These synthetic measurements demonstrate scaling rather than a general contract speedup.

Written with GPT 6.1 Sol on Extra High.

[tvolk131]

Draft reminder: @tvolk131 needs to hand-review this change before undrafting it (marking it ready for review). Please keep this PR in draft until that hand review is complete.