v0.2.0

The observability release. A GPU deployment that silently ran everything on
the CPU used to produce identical rows, identical EXPLAIN tags and nothing
above debug — this release makes that visible, from a counter in the
dashboard to a Prometheus endpoint on every worker.

It does not complete the production-readiness milestone. Ten lines of
Phase 8 remain open; three of them (#28, #45, #46) need a
CUDA device to verify on and one (#29) is what the spill manager's real
wiring waits for.

Added

• The CPU-fallback counter (#32). fallback_batches on OperatorStats
  and OperatorSnapshot counts every batch that took the CPU reference path
  although its operator was placed on a device. The TUI Inspector shows
  cpu fallback N / M — green at zero, yellow when some batches fell back,
  red when all of them did — and each operator logs one warn! on its first
  fallback naming the reason. A silent fallback was previously detectable
  only by re-running with RUST_LOG=oxidelake_compute=debug.

• Placement notes in oxide explain (#32). A node the rule leaves on the
  CPU is an ordinary DataFusion operator in the plan, indistinguishable from
  one that was never eligible. oxide explain now prints, under the plan,
  every node that was skipped and why — AggregateExec: \avg` is not one of
  the aggregates the device kernels implement (sum, count, min, max), the
  grouping key `s` is Utf8View; the device kernel groups on Int64`. A plan
  that lowered completely prints nothing.

• Spans, a per-query log line and a metrics endpoint (#33). An
  oxide.operator span per operator and partition carrying operator,
  partition, target (what the plan says) and backend (what is
  executing). OxideSession::collect logs one INFO line per query with the
  mode, rows, elapsed time and fallback count; oxide sql uses it, so
  RUST_LOG=oxidelake_runtime=info oxide sql … prints one line per query.
  Behind the new metrics feature, oxide-worker --metrics-port and
  oxide-scheduler --metrics-port serve oxide_* counters as Prometheus
  text. Cluster executors never run the placement rule, so the plan codec
  attaches a process-wide TelemetryHub to every Gpu*Exec it decodes —
  without that a worker's /metrics would describe no work at all. The
  endpoint is unauthenticated, like the Ballista ports themselves: bind it on
  a private interface. Without the feature --metrics-port is refused, never
  ignored.

• --batch-size, --output table|json|csv, worker --backend (#49).
  --batch-size overrides the CPU default (8192) and the GPU-target default
  (65536) on sql, explain and tui; zero is refused rather than silently
  returning nothing. --output json writes an array of objects and
  --output csv RFC 4180 with a header — an empty result still prints the
  header and [], so a script can tell "no rows" from "the query failed".
  oxide-worker --backend mirrors OXIDE_BACKEND and is applied before the
  first task, so it reaches the operators rather than being accepted and
  ignored. Every flag and environment variable is now in one README table.

• SessionOptions (#49), with OxideSession::local_with_options and
  connect_with_options. #[non_exhaustive] with builder methods, so a knob
  added later is a minor release.

Changed

• predict reads its activation from the model file (#47). Breaking
  for existing model files. A safetensors header declares
  {"__metadata__": {"oxidelake.activation": "relu"}} — relu, gelu,
  sigmoid, tanh or none, applied between the Linear layers and never
  after the last. A file that declares nothing is refused with a message
  naming the key and the alternative: predict(path, features, 'relu'), a
  new third argument that must agree with the header when the file has one.

  Fixing this found a worse bug. The loader built a stack of bare Linear
  layers and applied no activation at all, while the module docs, the
  README and ADR-0015 all said "ReLU between them" — so every non-linear
  model returned a linear model's numbers, and the only test of the numbers
  compared them against the same bare stack. The none case is now checked
  against oxmera's own Sequential::forward, which does not share the
  loader's loop, and relu, gelu and none must disagree with each other.

• Public enums are #[non_exhaustive] (#41). BackendKind,
  Compression, SessionMode, MemoryClass, MemoryTier, Tier,
  ModelSpec, Activation and the TUI's Panel / KeyInput / Transition
  now need a _ arm in a downstream match. ModelSpec::SequentialMlpRelu
  is gone, replaced by ModelSpec::SequentialMlp { activation } (see
  predict, below). The plan vocabulary stays
  exhaustive on purpose — Predicate, Comparison, Literal,
  AggregateFunction, DistanceMetric and the codec's GpuNode: every
  kernel must handle every variant, so a _ arm in the CPU correctness
  reference is a wrong answer that compiles. docs/RELEASING.md records the
  distinction, and the DataFusion coupling that cargo-semver-checks cannot
  see.

• The plan codec carries a version fingerprint (#42). postcard is not
  self-describing, so a GpuNode field added or reordered between builds
  decodes without complaint into the wrong parameters and a mixed-build fleet
  computes a confident wrong answer. Every encoded plan now carries a magic,
  a format version and a fingerprint over the crate version and the compute
  layer's feature mask; a plan from a different build is refused at decode
  with both fingerprints named. A committed snapshot pins the bytes of all
  four variants, so a wire-format change cannot land unnoticed.

• Memory-tier gauges show what the backend reports (#25). Capacities come
  from the local backend's MemoryInfo — the device's total for a GPU
  backend, host RAM for the CPU one — instead of the fixed 8/4/2 GiB
  constants the gauges used to be drawn against, and a tier the engine has no
  number for is drawn without a ratio rather than against a plausible one.

• The dashboard says the spill manager is library-only (#25).
  SpillManager is not on the query path: no operator registers a batch with
  it, so the tier gauges and spill counters stay at zero however much memory
  a query used. The Telemetry panel says so (spill library only: no query registers batches), as do the README, STATUS.md and
  docs/architecture.md. Wiring it in is not a matter of calling it from the
  join — the build side is probed by every batch — and waits on the streaming
  aggregate (#29).

• OperatorSnapshot and TelemetrySnapshot gained public fields
  (fallback_batches, capacity). Both are exhaustively constructible, so a
  downstream struct literal over them no longer compiles. They are snapshot
  types read field by field in practice, which is why they stay plain structs
  rather than becoming #[non_exhaustive] — a reader should be able to
  destructure one.

• An unavailable backend names the backend, not the way it was asked for.
  oxide-worker --backend cuda reaches the same code as OXIDE_BACKEND, so
  the message no longer sends the reader to a variable they did not set.

• Logs are uncoloured when stderr is not a terminal. ANSI escapes in a
  redirected log file are noise, and they break a grep for field=value.

Fixed

• oxide sql prints the columns of an empty result. The CSV header and
  the JSON [] are written from the plan's schema, which an empty result
  still has.