fix(vm): widen OP_LINE operand to 32-bit — line numbers wrapped at 65536 (#630)

[InauguralPhysicist]

Closes #630

The lexer, parser, and EigsChunk.lines[] all carry line as int and were correct. But the OP_LINE bytecode operand was 16-bit, so every line the VM tracked at runtime was line % 65536.

The sharp form: rc=0, wrong answer, no diagnostic

Two assignments exactly 65536 lines apart collapse onto one line identity:

what is x at 4464   ->  222      # should be 111

222 is the value bound on line 70000, which wrapped onto stamp 4464. No error, exit 0. Everything reading g_vm.current_line was affected — temporal at/when/state_at, runtime-error and stack-trace line numbers, and the trace-tape line context — all silently wrong past line 65535. That is exactly the regime generated programs (iLambdaAi) reach, and it's the one consumer that can't eyeball a wrong line.

Fix

OP_LINE carries a 32-bit operand (a program would need 2³¹ lines to overflow that). Line stamps are deduped — emitted only when the line changes — so the 2 extra bytes per change are negligible, and this is the correctness-first VM tier, not a perf path.

• No opcode number changed → the test_opcode_abi.c static-asserts are unaffected.
• No bytecode is serialized to disk (the tape records nondet events and re-compiles from source; --bundle attaches source, not bytecode) → no tape/bundle format bump.

Every operand-stride site (a missed one desyncs the walk silently)

site	change
compiler.c	chunk_emit_u32 / emit_op_u32 / emit_line emits 4 bytes
vm.c	read_u32 + CASE(LINE) + leaf-accessor mini-executor stride 4
chunk.c verifier	chunk_verify special-cases OP_LINE (4-byte skip); removed from op_verify_operands
chunk.c disasm	dedicated 32-bit branch; removed from op_has_u16
chunk.c	leaf-accessor compile-scan strides 4
jit.c	supported-prefix scanner + thunk emitter, 3→5 byte stride

Verification

• Repro fixed both tiers. The issue's exact case gives 111, 222 under interpreter and JIT (was 222, 222). A hot loop forced to JIT-compile past line 65536 returns the correct result and agrees byte-for-byte with EIGS_JIT_OFF.
• Error lines past 65535 now report the true line, not the wrapped one.
• Regression test [120] (generated at test time, like the [114] constant-pool-ceiling sibling): asserts the temporal query = 111/222, error line > 65535, and JIT/interpreter agreement. Each assertion fails on the pre-fix VM (222/222, line 4469).
• Verifier path (test_vm_run_bytecode, hand-built bytecode) green under ASan — confirms the chunk_verify stride stays in lockstep.

Gates

• Release suite 3078/3078
• ASan + UBSan detect_leaks=1 3082/3082, leak tally 0, no sanitizer errors
• make jit-smoke passes

🤖 Generated with Claude Code