lean_vm: add source-level failure stack traces

[latifkasuli]

Summary

• add source-level call-site metadata to compiled bytecode so runtime failures can reconstruct caller/callee frames from the VM frame chain
• replace history-based stack guessing with a bounded, cycle-safe fp unwinder that prints frames from the failing function back to main
• add ExecutionOptions / CompileAndRunOptions plus LEANVM_STACK_TRACE=0 for legacy callers to disable stack trace printing
• preserve worker-local pc/fp for parallel_range failures and hide compiler-generated loop helper frames from the rendered stack

Closes #112.

Root cause

The existing diagnostic path inferred call stacks from source-location history. That could show useful source context, but it was not organized by the live caller/callee relation and could not reliably print the current VM frame chain from the failing function back to main.

Implementation notes

Touched surfaces:

• crates/lean_compiler/src/b_compile_intermediate.rs
• crates/lean_compiler/src/c_compile_final.rs
• crates/lean_compiler/src/ir/instruction.rs
• crates/lean_compiler/src/lib.rs
• crates/lean_compiler/tests/test_compiler.rs
• crates/lean_vm/src/diagnostics/stack_trace.rs
• crates/lean_vm/src/execution/runner.rs
• crates/lean_vm/src/isa/bytecode.rs

The unwinder follows the VM call convention (mem[fp] = return_pc, mem[fp + 1] = saved_fp) and has three safety guards: a 256-frame cap, repeated-fp detection, and checked memory reads. If the cap is hit while another frame is readable, the rendered trace includes a truncation marker.

Bytecode gains debug-only call-site metadata in call_sites_by_return_pc. I checked repository serialization usage: Bytecode does not derive or implement serde serialization/deserialization, and aggregation bytecode is compiled into a OnceLock at runtime rather than loaded from a persisted bytecode blob.

Sequencing / collision notes

This intentionally opens as a draft because the collision surface is real:

• Custom thread pool #239 touches crates/lean_vm/src/execution/runner.rs and crates/lean_compiler/src/c_compile_final.rs, including parallel execution work.
• Goldilocks #210 touches runner.rs, c_compile_final.rs, and test_compiler.rs.
• lean_compiler: avoid duplicate compile-and-run output #245 overlaps crates/lean_compiler/src/lib.rs; if lean_compiler: avoid duplicate compile-and-run output #245 lands first, this should be rebased before review.

Validation

• cargo test -p lean_compiler runtime_failure -- --nocapture
• cargo test -p lean_compiler
• cargo fmt --check
• git diff --check
• cargo check --workspace
• cargo clippy --workspace --all-targets -- -D warnings