[DanielLuning]

What changed

Compiler time gets a permanent, reliable harness, and the symbolic-bounds chain is measured again against its +15% ceiling (ADR-0007).

• The harness: DEFK_COMPILE_TIME=<file> (read in Defk.__before_compile__, timed_compile/2). Every module compiles once for its result and once more in a fresh process; that second compile is measured and its microseconds and reductions are appended to the file. Without the variable, nothing changes, as with DEFK_DUMP_C.
• dev/compile_time.exs: compares interleaved rounds of two checkouts on the modules both have, by time or with --reductions, with and without SatEx.Kernels.
• Docs:
  • docs/agents/runs.md: the harness replaces the temporary-patch method.
  • docs/performance.md: a new entry 6.38.
  • ADR-0007's cost note: the new numbers.

Results

Why the old method was wrong:

• It timed the first compile. Under mix dump_c, lib and test/support compile in parallel, so the first kernel module waited about 1 s for the compiler's own modules: of a 2.5 s round, about 1.0 s was that wait.
• Timing a second compile in the same process was no better: SatEx.Kernels took either about 240 or about 510 ms on the same code.
• With the fresh process, five compiles of SatEx.Kernels agree to ±2% in time and ±0.05% in reductions.

Measured again (99 shared modules, two interleaved rounds each, medians):

From → to	Reductions	Time
before step 1 (23e27970) → after step 4 (a371dcab): the chain	+20.8%	+13.4%
after step 4 → main (b90ed0ff): the x3a and rt2 refactors	+7.3%	+13.5%
before step 1 → main	+29.5%	+28.8%

So the compiler is about twice as far past its state before step 1 as the ceiling allows. The copy-analysis cut (defk-o08.3, the next PR) takes 9% of this back. Provisional: the laptop was on battery, with no quiet-machine check.

How it was checked

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

• The baseline and middle checkouts ran with the same harness as throwaway patches. The baseline also had the DEFK_NO_CC tooling commit (cbfb5725) applied, so that no round ran cc.

• Touches a promised surface? No: DEFK_COMPILE_TIME is a dev knob like DEFK_DUMP_C, not documented for users. No CHANGELOG.md entry.

• Touches test/stability/? No.