feat!: v0.2.0 — observability, declared breaks and the codec fingerprint

[vyncint]

v0.2.0 — the observability release

A GPU deployment that silently ran everything on the CPU produced identical
rows, identical EXPLAIN tags and nothing above debug. This release makes
that visible, and declares the breaking changes 0.2.0 exists to carry.

It does not finish the production-readiness milestone. Ten Phase 8 lines
remain open; #28, #45 and #46 need a CUDA device to verify on, and #29's
streaming aggregate is what #25's real wiring waits for. docs/roadmap.md
and the CHANGELOG say so in those words.

What lands

• CPU-fallback counter and placement notes (Count and surface per-batch CPU fallbacks and planner skip reasons #32). fallback_batches on
  OperatorStats/OperatorSnapshot, a colour-coded cpu fallback N / M
  line in the TUI Inspector, one warn! per operator on first fallback with
  the reason, and placement notes printed under oxide explain naming
  every node the rule left on the CPU and why.
• Spans, a per-query log line, a metrics endpoint (Add tracing spans on the query path and an exportable metrics surface #33). An
  oxide.operator span per operator and partition (operator, partition,
  target, backend); OxideSession::collect logs one INFO line per
  query; --metrics-port on the worker and scheduler behind the new
  metrics feature serves Prometheus text. The plan codec attaches a
  process-wide TelemetryHub to every node it decodes — without it a
  worker's /metrics would describe no work at all.
• CLI knobs (Add --batch-size, --output format and a worker --backend flag #49): --batch-size, --output table|json|csv,
  oxide-worker --backend, and one README table listing every flag and
  environment variable.
• predict reads its activation from the model file (predict: require an explicit activation in the model header instead of assuming ReLU #47). Fixing this
  found that the loader applied no activation at all while the module
  docs, the README and ADR-0015 all said "ReLU between them" — every
  non-linear model was scored as a linear one, and the only test of the
  numbers compared them against the same bare stack.
• Plan codec fingerprint (Version-fingerprint the plan codec and fail fast on mixed cluster builds #42) and non_exhaustive enums with the
  semver policy written down (Mark public enums non_exhaustive and write the DataFusion-coupling semver policy #41), with the plan vocabulary left
  exhaustive on purpose.
• Spill manager scope settled honestly (Connect SpillManager and MemoryInfo to the operators or mark them library-only in docs and the TUI #25): tier capacities now come
  from the backend's MemoryInfo, and the panel and docs say the spill
  manager is a library no query path calls.

Verification

Full gate locally: cargo fmt --check, clippy --workspace --all-targets --all-features -D warnings, cargo test --workspace and
--all-features (37 result lines, 0 failures), RUSTDOCFLAGS="-D warnings" cargo doc, cargo deny check, crate metadata, skill version, DCO,
no-AI-attribution.

cargo semver-checks --workspace ran locally before tagging this time —
skipping it is what failed the v0.1.4 tag. It passes for all nine crates at
0.1.4 → 0.2.0, and with --release-type minor forced it lists exactly the
breaks the CHANGELOG declares: ModelSpec's variant, the non_exhaustive
enums in core/memory/runtime/storage/tui/compute, and the two new public
fields on OperatorSnapshot and TelemetrySnapshot.

End-to-end, on a real spawned cluster: a worker's /metrics was empty before
a query and carried oxide_operator_{rows_in,batches,fallback_batches}_total
after it, with the oxide.operator span wrapping the fallback warning.

Closes #25
Closes #32
Closes #33
Closes #41
Closes #42
Closes #47
Closes #49