awf v0.3.0

A workflow-format release. It renames several author-facing keys (migrate per
the BREAKING list below), adds container-less run: steps and native stall
detection with retry-as-continue, and tightens validation. The workflow-format
version (version: 1) is unchanged, but existing workflows must apply the key
migrations to keep validating.

BREAKING (v0.3.0 format)

• reduce's quorum over: is renamed to field:. The per-branch boolean
  field a quorum counts is now declared as reduce: { quorum: <k>, field: <field> }.
  Migration: rename over: to field: under reduce.quorum only — a map's
  own over: (the fan-out worklist) is unrelated and unchanged.
• The top-level workflow input: schema is renamed to input_schema:.
  Migration: rename the workflow's top-level input: key to input_schema: —
  {{ input.<field> }} template references and a call step's own input:
  (the instance binding passed to a subworkflow) are unrelated and unchanged.
• prune's keep: top(<k>) is now keep: <k>. The function-call-shaped
  wrapper is removed since top-k was always the only mode. Migration: replace
  keep: top(<k>) with the plain integer keep: <k>.
• A signal where: clause now requires the {{ }} envelope, and correlates
  against the payload via a signal.<field> root. The old bare-identifier
  form (an implicit substitute-into-string-then-parse against the payload,
  with a string-quoting hazard) is removed. Migration: wrap the clause in
  {{ }} and prefix payload fields with signal. — e.g.
  where: 'candidate_id == "{{ hyp.id }}"' becomes
  where: "{{ signal.candidate_id == hyp.id }}". Every other root (input.*,
  step.*, run.*, an as: binding, …) resolves against the surrounding
  engine scope exactly as in if/loop/gate conditions, and the value is
  compared as typed data — no more quoting workaround for strings.

Added

• output_artifact is no longer containerless-only. Any agent step that
  declares output_schema — container-backed or containerless — may now also
  declare output_artifact: <name> to publish its typed output as a
  content-addressed artifact. The requires-output_schema and
  mutually-exclusive-with-output_files rules (AWF3014) are unchanged.
• map's over: accepts a literal YAML sequence. In addition to a
  {{ }} expression evaluated at runtime, over: may now be an author-fixed
  literal sequence (e.g. over: [a, b, c]) — a static, digest-pinned
  parameter sweep known before the run, as opposed to a runtime-sized
  worklist.
• Container-less run: steps (bare shell). A run: code step may now omit
  container:; it executes host-side under the native sandbox (auto-selected
  when the workflow declares no image), so a hello-world needs no image or
  digest procurement. A bare run: is rejected under --backend docker
  (AWF1065) — declare a container: or run native.
• map's concurrency: is optional (default 1, serial). Omitting it runs
  the map serially; an explicit value <= 0 is now rejected (AWF1012).
• Native stall detection + retry-as-continue. An idle watchdog cancels an
  agent step that goes silent longer than timeout.idle (distinct from the
  wall-clock timeout), turning a wedged agent into a retryable failure. The
  openai/codex live adapter forwards codex's reasoning-summary heartbeat and
  gets a generous default idle (~300s); every other adapter is opt-in, and
  AWF3016 warns when idle: is set on an adapter that surfaces no liveness
  signal. On a stall, a persistent-session step resumes the same conversation
  thread — retry: { recovery: continue }, the default for session adapters —
  instead of restarting from scratch; recovery: restart is the escape hatch
  (AWF1064 rejects any other value at validation time).
• awf validate strictly rejects unknown workflow/step keys (AWF1062). A
  stray or typo'd key anywhere in a workflow document — previously silently
  tolerated — is now a hard validation error.
• awf validate strictly rejects bare-integer durations (AWF1063). A
  timeout or retry.initial/retry.max value must be a quoted duration
  string (e.g. "300s", not 300) — a bare integer previously parsed as
  nanoseconds with no error, causing the step to time out instantly.

Changed

• Breaking (native backend): run workdirs moved from work/ to work/<run-id>/
  to isolate concurrent native runs. A native run started before this change
  cannot be resumed after upgrading (its workdir path moved). Docker runs are
  unaffected (already run-id namespaced). Start affected native runs fresh.