[Feature] Reduction-aware inlining

[katsujukou]

What & why

Makes the NbE reducer (Semantics, ADR 0020) share work instead of recomputing it, and bounds the one case sharing can't help (genuine code-size blow-up). Together these let the optimized (opt-ON) self-compile of PursWasm.CLI.Main complete — it previously hung indefinitely at Optimize.Specialize.

The hang had two independent causes, both fixed here:

1. NbE recomputation (exponential). eval re-evaluated each inline-set binding per use site (M1), and quote re-evaluated shared values per path (M2) — Θ(2^d) on a diamond-shaped inline DAG.
2. Code-size blow-up at Specialize (module 276). Two sub-causes: the specialization dedup key was the show-serialized lambda body (multi-KB strings built + retained as Map keys), and NbE fully inlined the derived genericShow dictionary of the large IR ADTs into multi-million-node normal forms.

What's included

ADR 0035 — sharing the NbE reducer:

• Layer A (b12f633) — memoize each inline-set binding's eval (Data.Lazy keyed by binding name). Kills the M1 re-eval. Byte-identical output.
• Layer B (d1d26f8) — SShared constructor: evalVar tags inline unfolds; unShared strips it at every consumption site so it never blocks a redex; quote CSEs shared values into a hoisted let. Kills the M2 re-quote. Diamond DAGs now normalize in O(d).
• Layer C lite (cc11c18) — DictElim.normalFormSizeCap: when a declaration's reduced form exceeds the cap, re-reduce with the inline context emptied (the binding stays an ordinary call — the --no-opt-correct shape). Bounds the genericShow-into-show blow-up. Plus Specialize.canonicalKey now hashes the canonical form (Serialize.Hash.hashString) and abstracts free vars in a single substMany pass instead of show-ing it per free var — removing the giant-string Map keys.

Tests:

• Test.NbeStress — a synthetic diamond-DAG guard (depth-20 normalizes to linear size, fails loudly if the exponential returns) + two Layer-C cap guards (fallback above the cap; still inlines below it). Wired into Test.Unit.Compiler.

Docs:

• ADR 0035 promoted to Accepted (Layers A+B+C-lite), ADR 0020 progress note, README index.
• ADR 0037 (new, Accepted as a design record — not implemented): the per-module-codegen / separate-compilation direction explored this cycle (singleton GC rec-groups, boxed module boundary, .pmi/.pmo-as-ANF, wasm-merge link). Included here as the recorded rationale; no code change behind it.

Verification

• Optimized self-compile of PursWasm.CLI.Main completes — writes a valid 8 MB index.wasm (previously hung at Optimize.Specialize).
• Unit 162/162, e2e 150/150, bin 12/12 (incl. showShadow/intGenericShadow/examplesRun — the genericShow path).
• Bench: no regression (some benches faster from Layer-B CSE sharing; rest within noise; baseline untouched). The Layer-C cap doesn't trigger on the bench corpus (no decl approaches the cap).
• examples/metatheory compiles and runs.

Byte-equality is not required (the current compiler isn't a verified reference); the gate is test-suite + bench + examples no-regression.

Checklist

CI passing is enforced by the required ci-gate status check, not by a box here.
The items below are the human-judgment gates:

• Docs updated — any behaviour/feature/representation change is reflected in docs/ and relevant ADR(s).
• Tests — added/updated for this change; a bug fix includes a regression guard in the routinely-run lane (unit / e2e), not only a slow test:bin script.
• Design changes have an ADR — added or updated, with its Status set (ADR-first).
• No perf regression — for optimizer / runtime / lowering changes, benchmarks compared (node bench/run.mjs <dir>; do not overwrite bench/snapshots/baseline.json). N/A otherwise.
  - [ ] Runtime GC-type / ABI / canonicalization — changes to the value-type substrate or the host/runtime ABI keep cross-module type canonicalization intact. N/A otherwise.