The k-framework-binary (private, via kup) and k-framework (public, via
cachix) publishes built the same three derivations on separate runners,
recompiling K four times (2 jobs x 2 OS) for work that needs two builds.
Fold them into one matrix job: the first publish populates the local Nix
store so the second's nix build is an instant store hit. The two
publishes stay independent (continue-on-error + if: always()) so a flaky
upload to one cache neither blocks nor masks the other, and a final
guard re-fails the job if either push failed, preserving the release
gate the two separate jobs provided. Drop the removed
cachix-release-dependencies from the release job's needs.

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…4929)

Co-authored-by: devops <devops@runtimeverification.com>

~Blocked on:
runtimeverification/haskell-backend#4151
~Blocked on: runtimeverification#4929

Lets pyk request the haskell-backend's per-request log bundle and write
it to disk, so downstream can diagnose fallback behaviour per RPC. The
request carries the list of log entries to capture and the response
returns them in-band; pyk owns the default set and writes each request's
bundle to its own file. This pairs with the haskell-backend
per-request-rpc branch, which adds the haskell-logging request field
([String]) and the haskell-log-entries response field to the five
JSON-RPC endpoints and captures the named entries independently of the
server's -l/--log-file configuration.

Changes:
- BoosterServer now emits the booster-form haskell-log CLI (--log-file
FILE, repeated -l ENTRY) via a new overridable _haskell_log_cli_args
hook; it previously sent the kore-rpc form (--log, --log-entries A,B),
which made kore-rpc-booster exit on a usage error before opening its RPC
port.
- KoreClient.{execute,implies,simplify,get_model,add_module} take
haskell_logging: Iterable[str] | None — the entries to capture for that
request — and send it as the haskell-logging wire field (omitted when
None); the optional haskell-log-entries bundle is parsed onto every
result type, and simplify returns a SimplifyResult (state, logs,
haskell_log_entries).
- KoreClient.last_request_id exposes the JSON-RPC id of the most recent
request issued by the client, used to name each per-request log file.
- CTermSymbolic owns the canonical default entry set
(HASKELL_LOGGING_ENTRIES), overridable per instance (and through the
cterm_symbolic factory) so downstream can tailor it; the on/off flag is
threaded through execute/simplify/kast_simplify/implies and
KCFGExplore.extend_cterm.
- CTermSymbolic takes a haskell_log_dir: when set, every RPC requests
the bundle and the entries returned on the response are written to
<haskell_log_dir>/<request_id>.jsonl (one JSON value per line, named by
request id so concurrent proof workers never collide).
- pyk prove --haskell-logging points that sink at
<save-directory>/haskell-logs/ (requires --save-directory).

Validation:
- make -C pyk check clean.
- Confirmed against the per-request-rpc backend implementation: it
accepts haskell-logging: [String], routes the flat list to both engines
(skipping names each does not recognise), tag-matches id-carrying
contexts like Rewrite, and returns the captured entries in-band — pyk's
default set is fully recognised, no skips.
- Integration and regression-new not run locally (the machine's K
install is pinned for a benchmark) — relying on CI.

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>