[DanielLuning]

What changed

The spike defk-rt2.11 asked whether the compiler can hand Elixir mode its node types, so that Elixir mode only emits. The owner decided to close it and keep the typing twice. This PR records that decision and removes a dead write.

• ADR-0008: a dated extension, "the rest of the typing also stays twice". It records:

  • the measured seam;
  • the decision and its two reasons;
  • a "How to revisit" note.

  The status line names the extension.

• The dead write: Defk.ElixirMode.definitions/3 set the compiler ctx's :file "so err! reads it". Nothing on Elixir mode's path reads the ctx: a rejection there raises a plain ArgumentError (raise_form!/4), and none of the helpers it calls touches the ctx. The write and the unused file argument are gone, so the function is now definitions/2. Its body uses an implicit try.

• Arity references: updated in the caller (lib/defk.ex), test/compiler_ctx_test.exs, the compiler map and the inventory row of ADR-0006. No file under test/stability/ or test/contracts/ names the function.

Results

The spike, measured on every test and example module with a throwaway probe (lower_expr/2 against expr/2, through mix dump_c; the full numbers are in the ticket comment):

• The key: a unique id stamped into each node's meta after the desugar step keys a node. Line and column cannot serve: about 58,000 typings of macro-built kernels have no column. The stamp left the C of all 108 modules byte-identical.
• Agreement: 57,877 of 58,361 keyed Elixir-mode typings (99.2%) equal the compiler's own type.
• The rest, four special cases:
  • 71: [u8] against bitstring naming.
  • 35: typed twice by the compiler (the u64 retries, the in guard of case).
  • 92: float trees, which never pass through lower_expr.
  • 286: helpers, which the compiler only inlines.
• Why not: the u64 accumulator inference decides results (3 - 5 raises as u64, gives -2 as i64), so sharing it removes the agreement tests' check of that decision. Without it, about 20 lines remain to gain.

How it was checked

• mix ci on the laptop (macOS) passed: 2300 tests, Credo found no issues, Dialyzer reported 0 errors.

• Touches a promised surface? No: definitions is @doc false and hidden (ADR-0006). No CHANGELOG.md entry.

• Touches test/stability/? No.