[DanielLuning]

What changed

• bench/call_cost.exs and bench/call_cost_nif.c: the base cost of a kernel call, stage by stage, measured on the generated C itself. The bench module's C is varied with string replacements (no sigsetjmp, no fault-containment stores, no entry clock read, no slice init), each variant built and loaded as its own library, against a bare-NIF floor.
• priv/defk_runtime.h: the slice clock is defk_now_us. On Apple platforms it reads mach_absolute_time, converted with the timebase read once per library at load; other platforms keep enif_monotonic_time. The forced-yield test path still reads the ERTS clock once per forced yield (see below).
• Docs: docs/performance.md 5.13 holds the breakdown; 5.1, 4.9 and 8.2 point at it; docs/project.md has a dated line under the results snapshot; docs/beam-nifs.md names the clock. The four small-input rows are re-measured and stored (examples/scoreboard.json, README tables regenerated).

Results

Median of 9 interleaved rounds of 1M calls, minus the empty loop, M-series Mac, quiet (other processes at 0.1-0.8 cores):

call	before	after
Elixir remote call	0.2 ns
bare NIF / plain NIF add2 / add2 + sigsetjmp	3.7 / 6.2 / 7.7 ns
defk add/2 (no loop)	11.5 ns	10.7-11.0 ns
yielding kernel, one trip, NIF only	33.4 ns	12.4-12.6 ns
yielding kernel through its function (hint BIF)	37.3 ns	21.2-21.6 ns
enif_monotonic_time / clock_gettime / mach_absolute_time over a bare NIF	26.7 / 15.6 / 4.5 ns

The entry clock read was 22 ns of a tiny yielding call. A kernel without a loop adds about 5 ns to a plain NIF that decodes the same terms (the sigsetjmp 2.5, the fault-containment stores 1.2, the rest 1.6): nothing worth cutting there. The hint BIF (about 9 ns) stays: it is what spreads short calls (4.9).

Stored rows (quiet, 2026-10-19 -> now): ws_masking defk 1.231 -> 1.129 ms per 10,000 frames (123 -> 113 ns a frame; 1.46x -> 1.47x because the Elixir side also measured faster this session), geohash_crc one point per call 1.953 -> 1.657 ms (1.29x -> 1.58x), poor_tiny_calls 53 -> 45 ns a call (no loop, so nothing changed on its path: session spread), utf8_valid unchanged. bench/tick_gap.exs on one scheduler: yielding kernel p99 2.26 ms, worst 2.27-2.44 ms, the same as the snapshot's 2.25-2.3.

Found on the way: an ERTS clock read stores the time in the scheduler's data (erts_get_monotonic_time sets last_monotonic_time and resets check_time_reds), and the scheduler bumps its timer wheel from that stored time. Forced yields charge no reductions, so with the cheap clock alone a test process that slept 1 ms on the kernel's scheduler woke only when the kernel ended (bitstring_test, "a process killed mid-build" timed out). The forced path now reads the ERTS clock once per forced yield, as the old entry read did at every re-entry. A normal yield charges reductions through enif_consume_timeslice, so the scheduler reads its own clock there; the tick-gap numbers above confirm it.

How it was checked

mix test twice after the header changes (2305 passed; the one failure above found and fixed between them), mix ci on the M-series Mac (running at the time of opening; the result follows as a comment), bench/tick_gap.exs, and the four examples through mix defk.examples --bench --only ... --snapshot. The stored rows' stamps name the measured commit (0377111); the second commit is docs wording only.

• Touches a promised surface? The runtime contract's promise (a kernel yields, the slice length) is unchanged; the clock behind it is. A CHANGELOG.md entry says so under "no promise changes", added for visibility: drop it if the changelog is for promises only.
• Touches test/stability/? No.

[DanielLuning]

Local mix ci on the M-series Mac passed after the PR opened: 2305 passed (34 doctests, 906 properties, 1365 tests), every generated region up to date, exit 0.