v0.34.0 — the tasks envelope, honestly windowed

[0.34.0] — 2026-08-12

pve_tasks_list returns a windowed outcome envelope, and the repo now has exactly one
task-outcome classifier. The first deliberate brick of the domain layer: classify, don't
mirror. Named in public on the forum thread before it was built; every commit in this
release passed an independent adversarial review before tagging, three rounds, two of
which held the release back.

Added

• pve_tasks_list windowed envelope: {returned, by_outcome, tasks} with the lean
  field set (upid/type/id/user/status/starttime/endtime) and fields as the
  escape hatch (all = raw rows). by_outcome classifies each raw row server-side —
  running / ok / warnings / failed / unknown — by deterministic string matching
  on endtime + exitstatus text, never inference. Measured through the SDK's own
  call_tool on a live node: 6,676 → 5,589 wire tokens for 50 all-OK rows (16%; an
  all-OK window is the envelope's most favorable case — failed rows carry their full
  error text).
• There is deliberately NO total in this envelope. PVE truncates to the newest
  limit tasks before the server sees a row, so a full-history population count does not
  exist here — a count that only describes the fetched window must not wear the
  population's name. The tool description states this negatively (an all-ok by_outcome
  is never "no task ever failed") and routes "did anything fail" to errors=True, which
  live-provenly filters PVE's whole task history server-side and includes WARNINGS rows.
• Release gate regenerates lhm.plugin.json and fails on drift (before TOOLS.md,
  which derives from it): a skipped manual manifest regen could previously let both
  surfaces go stale together while the TOOLS.md drift check passed green.

Changed

• One task classifier repo-wide: pve_diagnose's failed_tasks now uses the same
  classify_task_outcome as the envelope. An error message that merely begins with the
  word "WARNINGS" now counts as failed on both surfaces (the exitstatus shape is
  WARNINGS: n); a statusless finished row classes unknown rather than failed; a
  garbage row fails closed into unknown, never open into a healthy-looking class.
• statusfilter descriptions teach the live vocabulary: ok / error / warning
  (each live-proven), and state negatively that by_outcome words (warnings,
  failed) and task-status words (running, stopped) are rejected by PVE with a 400.
  The old examples taught two values every call would 400 on.
• The health-check runbook prompt uses pve_tasks_list(errors=True) instead of "flag
  any that failed" over a 50-row window.

Fixed

• Three test fixtures modeled finished task rows without endtime — a shape live PVE
  does not produce (live-probed: 0 of 179 errored rows lacked it) — and one modeled
  errors=1 as a failures-only filter when live PVE includes WARNINGS rows. All are
  now live-faithful, and the tasks fixture honors limit/errors, so truncation can
  interact with the counts under test instead of being structurally invisible.