nx_opt_eqsat_pass.nx

nx_opt_eqsat_pass.nx -- EQUALITY SATURATION AS AN ACTUAL IR REWRITE PASS. This is the first thing in the estate that makes nx_eqsat CHANGE A PROGRAM rather than merely judge one. Everything before it -- seven gates, 93 teeth, and nx_opt_eqsat_oracle_gate -- proved the engine sound and used it to CHECK the shipping optimizer. This consumes it. WHAT IT DOES: for each two-operand integer instruction, build a small e-graph of that expression (constants as constants, non-constant operands as e-graph VARIABLES KEYED ON THEIR IR VALUE ID, so the same SSA value becomes the same e-class), saturate it under the data-driven rule table plus the const-fold analysis, extract the cost-minimal representative, and if that representative is a CONSTANT, fold the instruction's result to it. The rewrite mechanism is the incumbent one, copied deliberately from opt_const_fold: set the result Value's kind to CONST and its const_int. No new way to mutate IR is introduced by this file. WHY IT IS NOT A DUPLICATE OF opt_const_fold: that pass requires BOTH operands to be constants. This one folds through ALGEBRAIC IDENTITIES, so it reaches expressions the shipping pipeline cannot -- `x ^ x` and `x - x` with x a runtime value fold to 0 here and are left untouched there. That gap is real and currently unmitigated, because opt_simplify (the peephole that would catch some of these) is BISECT-DISABLED in opt_run and has been since a self-host miscompile. DELIBERATELY NOT IMPORTED BY nx_opt.nx YET, and the lang_opt_eqsat_pass watch contract stays OPEN because of it. Adding `import nx_eqsat.nx` to nx_opt pulls the whole engine into every nx_cc build, changing the compiler's closure and its bytes. That is a compiler change and it belongs behind the equivalence net (nx_cc_equiv_gate 10/10 + selfhost) as a deliberate act, not as a side effect of landing a pass. Shipping the pass first, wired to nothing, is the sequencing that keeps the shipping compiler byte-identical while the capability becomes real and provable. TYPE NARROWNESS, STATED: i32-typed instructions are SKIPPED. opt_const_fold wraps folded constants mod 2^32 to match the backend's masking behaviour, and a fold that ignored that would produce a silently wrong value -- exactly the class this estate tracks hardest. Rather than duplicate that wrapping logic (a second ruler), this pass declines i32 and says so. license_tier: ORIGINAL expect_exit: 0