Changelog

0.7.2

shackle:receive_response/2 — like receive_response/1 but with a millisecond timeout. Returns {error, timeout} if no reply arrives within the window. The arity-1 variant still blocks indefinitely (unchanged).
cast_error/0 is now an exported type — the sum no_server | pool_not_started | shackle_not_started of atoms that cast/2..4 can return.

cast/2..4 and call/2,3 error specs tightened from {error, atom()} / {error, term()} to {error, cast_error()}. No behavioural change; dialyzer now flags {error, typo} at call sites that don't match the sum. Existing no_server | pool_not_started | shackle_not_started were the only atoms ever returned.
Latency measurement in shackle_server switched from os:timestamp/0 + timer:now_diff/2 to erlang:monotonic_time(microsecond) + subtraction. The monotonic clock isn't NTP-jumpable and skips the kernel call, so latency telemetry now stays accurate across clock skew events and shaves ~50ns off per request.
The #cast.timestamp field type changed from erlang:timestamp/0 (the 3-tuple) to integer() (microsecond monotonic). The field is internal — clients that don't destructure the cast record see no impact. Custom shackle_client behaviours that read Cast#cast.timestamp receive an integer now.
vsn in shackle.app.src is now an explicit string ("0.7.2") — was git, which only resolved correctly when building from a checkout.

Bumped foil dependency from 0.1.3 to 0.1.4 (CI/infrastructure refresh; error/0 type tightened; macro-based DRY refactor with no perf change).
Bumped metal dependency from 0.1.1 to 0.1.2 (CI/infrastructure refresh; one system_continue/3 spec fix from ok to no_return()).

No source changes in shackle itself. Behaviour, API, and benchmarks are identical to 0.7.0.

Minimum supported Erlang/OTP version is now 25. OTP 21–24 are no longer tested and the compatibility code paths for those releases have been removed.

Replaced the granderl dependency with knot (in-house C NIF using wyrand). shackle_utils:random/1 now calls knot:uniform/1 — same signature, same range guarantees, no behavioural difference at call sites. granderl 0.1.5 on hex.pm fails to build on OTP 27+ (its preamble.sh invokes erl -noshell -s init stop -eval ... with options in an order that causes the node to terminate before -eval runs, leaving ERTS_INCLUDE_DIR empty); the fix exists in granderl's master but its hex package is not under our ownership, so we couldn't release a fix. knot is a tiny C NIF (~80 LOC) that builds correctly on every supported OTP and matches hand-written C dispatch overhead. Benchmarks (Apple Silicon, OTP 29, 10M iterations of uniform(254), median of 5 runs):
conc rand:uniform/1 granderl:uniform/1 knot:uniform/1
1 34 ns 13 ns 12 ns
8 9 ns 7 ns 3.3 ns
32 6 ns 8 ns 3.2 ns
128 6 ns 8.5 ns 3.1 ns
CI matrix now covers OTP 25, 26, 27, and 28; actions/checkout upgraded to v5.
telemetry dependency bumped from 1.2.1 to 1.4.2.
README documents the full set of telemetry events shackle emits and their measurement shapes.

conc	rand:uniform/1	granderl:uniform/1	knot:uniform/1
1	34 ns	13 ns	12 ns
8	9 ns	7 ns	3.3 ns
32	6 ns	8 ns	3.2 ns
128	6 ns	8.5 ns	3.1 ns

ATOMICS, DECENTRALIZED_COUNTERS, ETS_TAKE, and SSL_HANDSHAKE platform_define macros.
OTP_RELEASE ifdef block in shackle_internal.hrl and the associated pre-OTP-21 logger / stacktrace fallbacks.
shackle_utils:info_msg/3 and shackle_utils:warning_msg/3.
ETS-based fallback path in shackle_status (the persistent_term + counters path is now unconditional).

See git history for details.