#545)

* fix(cli)!: rename mm session start --json key stale_ended → auto_ended

The JSON list returned by `mm session start --json` carries UUIDs from
two distinct paths: cutoff-based stale cleanup driven by
`--auto-end-stale`, and the cross-agent forced-end inside `--idempotent`
when the active session belongs to a different `--agent-id`. The
previous key `stale_ended` only describes the first; cross-agent ends
are not stale, just superseded.

The DB metadata already disambiguates them via `metadata.reason`
(`stale` vs `cross_agent`) — this commit aligns the JSON key with the
metadata's existing `auto_ended` flag so the CLI surface and the
storage record use one vocabulary. Drops the misleading "stale" word
from the text-mode print line for the same reason. The shape only
shipped post-v0.1.32 (PR #543, b8a5673) and never went out in a tagged
release, so no released hook caller depends on the old key.

Followup item 2 from RFC `memtomem-docs#24` (per-entry `reason` shape:
flat vs split vs dict-enriched) stays deferred — that decision needs a
distribution measurement on `sessions.metadata.reason` after a few
weeks of real hook usage. The per-row `reason` is already on disk, so
no telemetry plumbing is required for that future analysis.

Co-Authored-By: Claude <[email protected]>

* fix(cli): show stale/cross-agent breakdown in text-mode auto-end count

Per PR review: ``mm session start`` text mode previously printed
``Auto-ended N session(s)`` after the rename, dropping the "stale"
hint that the prior wording carried. JSON consumers don't lose
anything (``sessions.metadata.reason`` carries the per-row split),
but a human running the command by hand can't tell whether the
auto-ends came from cutoff age or from cross-agent forced-ends.

Track per-reason counters at the two append sites — both code paths
already know which reason they're handling — and emit a
``(N stale, M cross-agent)`` suffix when either is non-zero. The
JSON shape (flat list of UUIDs) is untouched so the existing
``test_json_output_shape`` pin still holds.

New test ``test_text_mode_breakdown_by_reason`` exercises both
paths firing in the same call and pins the two count fragments,
since this is the only place the per-reason split is now surfaced
without going through SQL.

Co-Authored-By: Claude <[email protected]>

---------

Co-authored-by: pandas-studio <[email protected]>
Co-authored-by: Claude <[email protected]>