fix(kanban): a crash streak the card has left is historical, not a current crash - #110463

What this fixes

_rule_repeated_crashes (hermes_cli/kanban_diagnostics.py) counts a trailing crashed streak in which any non-completed/reclaimed outcome "neither counts nor breaks" it — so a newer run that ended unreported, blocked, spawn_failed or timed_out leaves the old streak intact and the finding then describes that streak in the present tense and scales its severity as though the newest run were one of the crashes.

Measured on the live vx board, 2026-09-14 01:16Z, hermes kanban diagnostics --severity critical --json (pre-fix):

unreported is the deliberate outcome for a run that exited cleanly without a terminal kanban call — so the probe was reading the conversion working, and paging critical, about crash runs two days old.

What changed

The streak is a current incident only while the task's newest run is itself one of its crashes:

• current (newest run crashed) — byte-identical to before: same title, same detail, same severity ladder (critical at 2× threshold, else error), so a real repeating crash and the pid <n> not alive mode page exactly as they did;
• finished — the finding keeps the streak at warning (below the critical floor the board-health page pages on) and says so, dated: Agent crashed 5x (historical, last 2026-09-12 17:29Z): … with a detail that names the newest run id/timestamp/outcome and states "so this streak is historical — not a current crash";
• basis — data now carries is_current, newest_outcome, newest_run_id, newest_run_ended_at, last_crash_run_id, last_crash_at and runs_read (the ordered list of runs the scan read, with their outcomes), so a reader can tell "currently crashing" from "crashed historically" without re-reading task_runs.

crash_threshold and the outcome vocabulary are untouched — this distinguishes, it does not suppress.

Scope mapping (the card's four jobs, and the review's criteria)

Evidence

RED — pre-fix source, new tests kept (git checkout <base> -- hermes_cli/kanban_diagnostics.py):

$ python -m pytest tests/hermes_cli/test_kanban_diagnostics.py -k repeated_crashes -q
FAILED …::test_repeated_crashes_finished_streak_is_historical_not_current
FAILED …::test_repeated_crashes_finished_streak_behind_a_blocked_run_is_historical
FAILED …::test_repeated_crashes_finished_streak_behind_a_live_run_is_historical
FAILED …::test_repeated_crashes_current_streak_pages_exactly_as_before
FAILED …::test_repeated_crashes_pid_not_alive_mode_still_pages
FAILED …::test_repeated_crashes_threshold_and_vocabulary_are_untouched
6 failed, 1 passed, 4 deselected
AssertionError: assert 'critical' == 'warning'

GREEN — fixed source: 11 passed (tests/hermes_cli/test_kanban_diagnostics.py).

Live board, same tree and same DB copy, pre-fix vs fixed (hermes kanban --board vx diagnostics --json, HERMES_KANBAN_DB pointed at a copy of ~/.hermes/kanban/boards/vx/kanban.db):

critical, pre-fix : t_8afe5089, t_6db2248b, t_2d283056, t_7a11b2a2  (4 repeated_crashes)
critical, fixed   : (none)
all severities, fixed:
  t_406e4b7e warning is_current=false newest=10573/unreported  Agent crashed 3x (historical, last 2026-09-11 03:53Z): pid 4079594 not alive
  t_8afe5089 warning is_current=false newest=10535/unreported  Agent crashed 5x (historical, last 2026-09-11 07:30Z): pid 364463 not alive
  t_6db2248b warning is_current=false newest=10524/blocked     Agent crashed 5x (historical, last 2026-09-11 17:41Z): …
  t_2d283056 warning is_current=false newest=10536/unreported  Agent crashed 5x (historical, last 2026-09-12 17:29Z): …
  t_7a11b2a2 warning is_current=false newest=9252/blocked      Agent crashed 8x (historical, last 2026-09-13 07:11Z): …

Rendered page — the board-health notifier (internal_processes/scripts/vx_board_health_notify.py --dry-run, VX_BOARD_HEALTH_DIAG_JSON fed the two diagnostics bases above) would send 4 repeated_crashes findings pre-fix and 0 fixed, on the identical board copy.

Regression suite (the 28 failures it reports are pre-existing — exonerated)

python -m pytest tests/hermes_cli -k kanban -q → 28 failed, 340 passed, 1 skipped, 11142 deselected in 878.47s. None of the 28 is in tests/hermes_cli/test_kanban_diagnostics.py (11/11 green above). Checked against a pristine worktree at this PR's base (ee4452991), splitting the 28 into two batches and running each batch on both trees:

14-ids batch: base 14 passed / fixed 14 passed      → diff of FAILED sets: identical (empty)
14-ids batch: base 14 failed / fixed 14 failed      → diff of FAILED sets: identical (empty)

So 14 reproduce on the untouched base tree, and the other 14 pass on both trees when run as that batch (order-dependent flakiness of this host's environment) — no failure is attributable to this change.

Preflight (implementer checks)

1. Selection identity — N/A: pure rule function of (task, events, runs, cfg); the run list is re-read and re-ordered per call (unchanged).
2. Coercion inventory — int(cfg.get("crash_threshold", 2)) unchanged. New _fmt_run_ts coerces ended_at or started_at with a phrase fallback instead of raising or printing None into an operator's page; a missing newest outcome reads "is still in flight (no outcome yet)". No write path.
3. Authz write surface — N/A: read-only, no endpoints, no permissions.
4. Success + visible errors — the new branch is explicit text plus an inspectable basis; the finding is not silently dropped (it stays in the CLI/dashboard at warning).
5. Off-path — N/A (no flag). The rule's exemptions (done/archived/running, unified-rule suppression) are untouched.
6. Tests assert something live — the new cases drive the real compute_task_diagnostics entry point with the board's own run shapes (ids, unreported/blocked, pid <n> not alive), and assert the rendered title/detail strings, not just severity.
7. Cross-repo API contract — the only consumer of this diagnostic is internal_processes/scripts/vx_board_health_notify.py: it pages --severity critical and fingerprints data minus its volatile keys. Additive data keys mean a one-time re-page per currently-critical crash finding (a changed fingerprint is a changed finding); the warning severity removes exactly the false pages. No consumer needs a companion change to keep working; the host carry is carded (below).
8. Backend staging & write safety — N/A: no external system is written; diagnostics are read-only and the live-board work above ran against a copy of the board DB.
9. Redirect / open param — N/A.
10. State + scope completeness — count == data["consecutive_crashes"], and the basis travels with the finding (runs_read); no hand-written duplicate of a derived value.
11. SQL reserved words — N/A: no SQL added.
12. Rate-limited endpoints / confirm-then-write races — N/A: no HTTP, no writes.

Adjacent observations (out of task scope)

• hermes_cli/kanban_db.py's task_runs.outcome DDL comment still lists only completed | blocked | crashed | timed_out | spawn_failed | gave_up | reclaimed; unreported is absent there because it is introduced by the still-open fix(kanban): park a clean exit with no terminal call, carrying its last-known state #110322, not by this change. Left alone deliberately: this PR must not depend on that one merging.
• The dispatcher host runs upstream + 18 local patches, so this PR alone does not change the page it renders. The carry is carded as kanban t_79c3fde9 (same route as t_ecf2866f / The-ValueExchange/internal_processes#492).