[DanielLuning]

Merge #68 first. This branch builds on defk-o08.2/compile-time (the timing harness it was measured with). Until #68 merges, this PR shows that commit too.

What changed

Copies.nested/3 asked three questions for every block nested in a statement: arm?/2, taken/2 and reads_outside/2. Together they made five walks of the whole statement, and taken/2 was quadratic, because it called holds_arm?/2 (a walk of the arms) on every node. This was the largest single cost in the compiler: about 10% of the time.

• Copies.arm_reads/2 gives the same answer in three linear walks:
  1. One prewalk counts the occurrences and checks that the node is directly an arm.
  2. cut/2 rebuilds only the path to the arm. It mirrors Macro.prewalk's descent, and on the way back it cuts each if, cond and case as other_arms_out/2 does.
  3. reads/1 reads what is left.
• The old functions stay as the spec. test/copies_test.exs has a property test that compares arm_reads/2 with arm? plus reads_outside(taken(...)) on every subterm of generated statements. The statements contain if, cond, case, matches and blocks, and they use few names, so that a node often appears twice.
• Docs: docs/performance.md 6.39, and ADR-0007's cost note.

Results

• The C: mix dump_c is byte-identical for all 108 modules. The only difference is the known path-hashed defk:loop key in Defk.ReportTest.Gen.c.
• Compiler work: measured with defk-o08.2: compiler time measured warm in a fresh process; the chain measured again against its budget #68's harness, two interleaved rounds against main plus the harness:
  • reductions: 138.6 M → 125.8 M (−9.2%; −5.0% without SatEx.Kernels);
  • time: 1,101 → 996 ms (−9.6%; −2.9% without SatEx.Kernels);
  • SatEx.Kernels: −8.4 M reductions, −86 ms.
• The budget: against the compiler before the symbolic-bounds chain, the work is now +17.6% (+8.4% without SatEx.Kernels), down from +29.5%. That is still above the 15% ceiling; two follow-up cuts are ticketed.

Provisional: the laptop was on battery, with no quiet-machine check.

How it was checked

• mix ci on the laptop (macOS) passed: 2302 tests, Credo found no issues, Dialyzer reported 0 errors.

• A mutation run: with the condition of an if kept instead of cut, the property test and the example test both failed. The line was then restored.

• The byte-identical mix dump_c gate, against an untouched main checkout.

• Touches a promised surface? No. No CHANGELOG.md entry.

• Touches test/stability/? No.