[M8] Implement evidence-backed compatibility-safe optimizations

[Teakowa]

Parent: #9

Depends on: #50, #51, #53

Planning status: provisional until the M7 gate is complete.

Goal

Implement only optimizations whose Workshop semantics and resource benefits are demonstrated by compatibility/scenario evidence and measured through Wright's shared cost model.

Scope

• Select candidate transformations from real generated WIR/resource profiles rather than hypothetical compiler theory.
• Implement correctness-first canonicalization and resource-reducing passes suitable for compat mode.
• Require N-level and relevant E-level regression coverage for each optimization.
• Keep transformations with unresolved Workshop typing/timing behavior disabled in compat and, if useful, experimental in aggressive.
• Record before/after WIR and [M8] Define Workshop cost model and performance/resource benchmarking #53 cost/resource snapshots for review/debugging.
• Reuse [M8] Define Workshop cost model and performance/resource benchmarking #53 cost semantics rather than defining pass-specific resource scores or incompatible metrics.

Non-goals

• Maximizing optimization count.
• Enabling aggressive transformations by default.
• Optimizing away observable Workshop timing/control-flow semantics.
• Inventing optimizer-local cost definitions that cannot be consumed by other Wright APIs.

Acceptance criteria

• Every compat optimization has evidence-backed semantic safety and measurable benefit or canonicalization value.
• Optimized programs pass the relevant [M8] Finalize v1 compatibility matrix and S/D/N/E release gates #49/[M8] Build runtime-sensitive semantic scenario framework #50 gates.
• Resource/benefit claims are expressed through the shared [M8] Define Workshop cost model and performance/resource benchmarking #53 cost model and are comparable across passes.
• Disabling optimization preserves a clean reference path for debugging.

[Teakowa]

Delivered in commit 2537396 (crates/wright-transform/src/fold_constants.rs, synthesize_initializers.rs).

• fold-constants: arithmetic on literals (2 * 3 → 6, matching the expressions-values oracle), unary negation, sqrt/abs folding (overpy-cake oracle folds sqrt chains to 16-digit values), comparisons on literals, boolean logic, valueInArray(x, 0) → firstOf(x), and vect(0, 1, 0) → Up (cake oracle evidence).
• synthesize-initializers: moves non-trivial variable initializers into the reference's synthetic 'Initialize global variables' rule (oracle evidence), clearing the variable tables.
• Each pass is evidence-backed with N-level regression fixtures (the corpus fixtures byte-compare against the oracle under compat: basic-rule and overpy-cake byte-identical; the rest pass modulo documented debug/print presentation) and before/after PassStats metrics in the pipeline tests.