v5.18.0

Changed

• Scheduler: runqueue latency no longer counts the idle task, so it measures
  time real tasks spent waiting instead of time CPUs spent asleep. Reported
  latency drops, and the spurious outliers in the top histogram bucket — an
  underflow from a slot shared across CPUs — are gone. Alert thresholds tuned
  against the old numbers will need revisiting. (#1035)
• Scheduler: involuntary context switches no longer count a CPU waking from
  idle as a preemption. On a lightly loaded machine this was most of the count,
  charged to the root cgroup, so systems appeared to be under preemption
  pressure that did not exist. (#1046)
• Scheduler: context switches are now split into voluntary and involuntary.
  Only involuntary was emitted before, which made "no voluntary switches"
  indistinguishable from "not measured" — so a blocking handoff between two
  processes on a pipe looked the same as real preemption, despite pointing at
  completely different causes. (#1040)
• CPU: a task's CPU time is no longer lost when it exits. The counter and its
  labels are now released together, so the series ends cleanly rather than the
  task's total going missing. Workloads with heavy process churn were
  undercounted. (#1037)
• BPF: perf counters that never measured anything are no longer published, so a
  zero can be trusted. A counter that never got a PMU slot used to be
  indistinguishable from one that legitimately measured zero, which is common
  under virtualization where the PMU may not be available at all. Metrics that
  were previously present-and-zero will now be absent on such hosts. (#1052)
• Agent: snapshot_format defaults to v3, which reshapes /metrics/json from a
  flat {counters, gauges, histograms} into {groups: [...]}, and changes
  /metrics/binary with it. In exchange, each reading carries the window it was
  actually acquired over, making rate() error bars precise rather than
  approximated — and scrapes cost roughly half the wall time and a third of the
  agent CPU. Anything parsing those endpoints directly needs updating; set
  [general] snapshot_format = "v2" to defer. (#1076)
• Recorder: rezolus record writes rezolus.rez by default rather than
  rezolus.parquet. The extension picks the format now, so --format is rarely
  needed. Scripts that record with no arguments and then read rezolus.parquet
  will not find it — they fail at the read rather than silently recording the
  wrong thing. A --format contradicting the extension is an error rather than
  a silent choice between them. (#1097)
• Recorder: record --node and --instance are gone. No reader was ever
  written for either, so they accepted a value and dropped it; anything that
  looked like it was labelling a recording was not. --metadata node=NAME and
  --metadata instance=NAME write the keys combine actually reads. Passing
  the old flags is now an error rather than a silent no-op. (#1097)
• Recording: the parquet subcommand group is now recording, since these
  commands stopped being parquet-specific once .rez arrived. rezolus parquet ... still works, so existing scripts keep running. (#1075)
• Recorder: the tar .rez container is no longer written. Archives from older
  releases still open everywhere, and recording upgrade converts one. (#1074)
• Status: rezolus status with no arguments asks the agent on this machine, and
  accepts a bare host or host:port. It exits non-zero when any sampler is
  unhealthy, so it works as a readiness probe or a gate in a script. (#1099)
• Agent: a cached snapshot is served without re-encoding it. Every request used
  to rebuild a multi-megabyte body and throw it away; 500 requests against one
  cached snapshot went from about 2 MB of RSS growth to none. (#1084)

Added

• Recorder: .rez recordings, where each sampler records at its own cadence and
  carries the window its reading actually covered. That window is what lets
  rate() and irate() report uncertainty bounds instead of a bare number, so
  you can tell a real change from sampling noise. Archives are around a third
  the size of the equivalent parquet and query about twice as fast. A recording
  is also valid at every instant: it is readable while still being written,
  survives a SIGKILL or power loss having lost at most one interval, and stops
  in the same bounded time whether it ran for a minute or a day. (#1017, #1041,
  #1060, #1061, #1070)
• Recording: queries spanning two samplers of a .rez are answered instead of
  refused, so ratios like cycles-per-instruction across samplers work. Where the
  two cadences differ materially the query is evaluated at the slow sampler's
  own timestamps, so its values are not silently held forward across points it
  never measured. (#1080, #1082)
• Recording: convert turns a raw msgpack recording into parquet after the
  fact, so a long unattended capture can record at the lowest possible overhead
  and pay the conversion cost later. (#1059)
• Recording: upgrade brings a v1/v2 (tar) .rez onto the current container.
  combine, filter and annotate do it in passing, so older archives keep
  working without a migration step. (#1073, #1074)
• Viewer: --tui explores a recording or a live agent from the terminal, for
  looking at a machine over ssh without forwarding a port or opening a browser.
  (#1019)
• MCP: extract-features summarizes a recording as a deterministic, versioned
  JSON record, so an agent can assess a capture without reading every series.
  (#1028, #1029, #1030, #1031, #1032)
• Agent: PMU hardware counters are budgeted rather than over-subscribed.
  Excess events used to open successfully, never get scheduled and read nothing,
  with which samplers won decided by an unspecified link order that could change
  between builds — and on a hypervisor this reached past the host, leaving
  guests counting nothing while reporting themselves healthy. Counters are now
  allocated per-PMU by rank, all-or-nothing, and rezolus status names the
  samplers that lost as pmu-limited or pmu-starved. reserved_pmu_counters and
  reserved_pmu_cpus leave headroom for other tooling on the box. (#1093,
  #1096)
• Viewer: rate time-alignment modes (Aligned/Raw) (#1023), sampling-jitter
  visualization for simple-capture parquets (#1014), and a swatch row explaining
  what each shade means on charts with band fills (#1021).

Fixed

• Agent: a config file that omits the [general] table starts with the
  documented defaults instead of exiting with "ttl couldn't be parsed: value was
  empty". (#1099)
• Agent: cpu_perf opens its per-CPU events as one perf event group, so its
  counters describe the same slice of time and derived ratios are meaningful.
  (#1091)
• Agent: a host with no GPU no longer ships GPU tables in every recording. It
  used to carry 448 phantom AMD members and 672 NVIDIA ones, all null. (#1081)
• Agent: groups with no live sampler declare no members (#1086), and V2 external
  metrics get their own column key (#1090).
• Recorder: a .rez recording that cannot start now exits non-zero. It printed
  an error and exited 0 before, so a record && analyze pipeline carried on and
  analyzed whatever a previous run had left at that path. A usage error also
  prints one line instead of a panic backtrace. (#1097)
• Recorder: a malformed histogram cell can no longer wedge a .rez recording
  (#1077), and segment size accounting is corrected so seals are not
  mistimed (#1050).
• Exporter: a histogram that cannot be downsampled is dropped rather than
  panicking, so one bad metric no longer takes the exporter down. (#1088)
• CLI: six help-text claims that contradicted the code were corrected — three
  commands said v3 .rez archives "error clearly" long after they worked, which
  told users a working operation was unavailable. A test now fails the build if
  a help text names a flag its command does not define, or if an example
  invocation uses one. (#1098)
• CLI: the published documentation was five weeks behind the binary, listing two
  flags that now exit non-zero and omitting status, .rez and --tui
  entirely. (#1100)
• Analysis: subsystems are inferred from metric names, and absences that cannot
  be known are never asserted. (#1034)
• Viewer: restored a missing boxplot.js symlink that had taken the deployed
  viewer down, with a CI guard against recurrence. (#1012)
• Installer: the apt repository architecture is detected instead of hardcoded to
  amd64, so arm64 installs work. (#1018)