Release v0.3.0

Added

• Self-describing results. A RunResult (and the persisted
  cases/<key>/result.json) now carries the sample's input (the prompt turns
  sent) and expected (the reference value, when the dataset provides one), so a
  saved result can be read back without the original dataset. Both are optional
  on the wire — input omitted when empty, expected when absent.
• Docs diagrams. Five new committed SVGs visualise the model, each in its
  topical guide: the end-to-end workflow (mira-workflow.svg — author →
  plan → execute → score → report) in getting-started.md,
  the entity hierarchy (mira-entities.svg — study ▸ eval ▸
  dataset/subject/scorers/targets/axes, expanded into cases · trials ·
  transcripts · scores) in authoring.md, the host ⇄
  study run lifecycle (mira-run-lifecycle.svg — the protocol sequence for one
  run) in how-it-works.md, the subject fan-in
  (mira-subjects.svg — the three subject shapes normalising into one
  Transcript) in subjects.md, and the scoring flow
  (mira-scoring.svg — transcript surfaces → scorers → case verdict) in
  scorers.md. Indexed in
  docs/README.md.
• JSONL and CSV report formats (--format jsonl / --format csv) for
  un-aggregated, analysis-ready exports. jsonl writes one RunResult per line
  (lossless — the line-delimited dual of json); csv is long-format, one row
  per (case × score) with the case columns repeated and open-vocabulary
  metrics/metadata flattened into stable metric.*/meta.* columns. Both
  work anywhere --out/--format do (run, report, score); a --group-by
  view is intentionally not folded in — the consumer aggregates the rows.
• Per-case wall-clock timeout: give up on a case after a budget of seconds,
  cancelling the in-flight run (best-effort cancel over the protocol) and
  recording it as a failed case. Set it on the CLI (mira run --timeout SECONDS,
  all targets), per target in mira.toml ([targets.LABEL].timeout), or as a
  preset default ([presets.NAME].timeout). Precedence, first set wins:
  --timeout > per-target > preset; unset ⇒ no limit. A timeout is non-retryable
  (retrying would burn the same budget) and counts as a target failure.
• Glob case selection. --targets, --samples (new), and --evals (new)
  match the target label / sample id / eval name by glob (*, ?, [set],
  {a,b}); a literal value stays an exact match. --axis values are globbed
  too. A small dep-free matcher (mira::glob_match) backs both the host and the
  in-process Runner (Runner::samples(…), glob-aware Runner::targets(…)).

Changed

• BREAKING (preset): the preset filter key is replaced by per-dimension
  samples (glob on sample id). targets/samples/evals in [presets.NAME]
  now glob-match and accept either a single string or a list. The cross-cutting
  case-key substring stays available as the positional mira run [filter].