fix(vm): pass raw error values through pcall and add §6.1 position prefix

[davydog187]

pcall/xpcall pass the raw Lua error value through; error() adds the §6.1 position prefix

Plan: .agents/plans/A48-pcall-error-value-passthrough.md
Closes #334

Goal

Per Lua 5.3 §6.1, error(value) raises an arbitrary Lua value and pcall returns it as-is as its second result; xpcall hands it to the message handler untouched. Previously extract_error_message/1 stringified every non-string value (error({code = 1}) → "table: 0x..."). Additionally, string messages now gain the reference source:line: position prefix (error("boom") → "file:1: boom"), suppressed by level == 0 — all without disturbing host-facing error rendering.

How

• lua_error captures the level argument and builds the prefixed view into a new Lua-facing-only :lua_value field on RuntimeError; :value stays raw, so message/to_map/stringify (and the (error object is a TYPE value) host formatting) are provably untouched — no doubled location.
• extract_error_message/1 → error_value/1: lua_value → raw value (matched on key presence, so error(nil)/error(false) pass through) → Exception.message fallback for ArgumentError/plain Elixir exceptions.
• Dispatcher parity: native calls inside compiled closures publish a {nil, source} position, so the prefix is suppressed rather than attributed to a stale outer line (the dispatcher strips per-instruction lines at encode time). The interpreter attributes correctly.

Success criteria

• New dual-engine matrix test/lua/vm/pcall_error_value_test.exs (26 cases, both engines): table/number/boolean/false/nil passthrough, prefix shape + source-first ordering, level-0 suppression, xpcall raw handler value, erroring-handler fallback to the original raw value, TypeError/bad-argument regression guards — all green (red-first, commit 424c65a)
• mix test: 2234 passed, 0 failures; host-message canaries unchanged (error_gallery_test.exs, integration_test.exs:1382, lua_test.exs:940, runtime_exception_test.exs:34-35)
• mix test --only lua53: 17 passed / 12 skipped — identical to main baseline
• mix format + mix compile --warnings-as-errors clean (mix credo is not a task in this repo — see Discoveries)

Changes

CHANGELOG.md                                   | 13 ++++
lib/lua/vm/executor.ex                         | 17 ++++-
lib/lua/vm/runtime_error.ex                    | 11 +++-
lib/lua/vm/stdlib.ex                           | 69 ++++++++++++++++-----
test/lua/vm/pcall_error_value_test.exs         | 285 +++++++++++++++++++++
test/lua/vm/pcall_state_preservation_test.exs  | 24 +++++--

Discoveries

• The position-suppression edit means native raises (assert, bad-argument) inside nested compiled closures now render a source-only host header (at <eval>:) where they previously had no location header at all ({nil, nil} position). Verified via before/after canary: strictly added information, no wrong line, no test pinned the old bytes; ErrorFormatter.format_location(source, nil) is an explicit supported clause.
• The plan's verification listed mix credo, but credo is not a dependency of this repo.
• Follow-up needed (issue not yet filed): a dispatcher native-call line bridge so error() inside compiled closures gets a real line (collapsing the engine-conditional prefix assertions) and error(msg, 2) caller attribution becomes implementable.

Out of scope (intentional)

• error() level >= 2 (needs per-frame line numbers; treated as level 1)
• Position prefix on internal raises (TypeError/ArgumentError — host ErrorFormatter already renders location; many tests pin exact strings)
• Correct compiled-closure prefix line under the dispatcher (suppressed instead; see Discoveries)
• Unskipping test/lua53_tests/errors.lua (blocked on an unrelated load() parse-error-prefix mismatch; its level-0 and error()-nil cases are reproduced in the new unit test)
• Any change to host-facing rendering

Verification

mix format                          # clean
mix compile --warnings-as-errors    # clean
mix test test/lua/vm/pcall_error_value_test.exs   # 26 passed
mix test                            # 2234 passed, 0 failures, 19 skipped
mix test --only lua53               # 17 passed, 12 skipped (== main baseline)

[davydog187]

Process dictionary is really a hack. I think this was from a previous design choice made for performance that we need to revisit

[davydog187]

we will evaluate this later