…NL on delete (#4868)

Channel `/new`, `/reboot`, `/compact` previously called
`kernel.reset_session(agent)` without a scope, which deleted EVERY session
the agent owned (default + every per-channel sid + cron-spawned ones).
Typing `/new` in Telegram silently nuked the dashboard transcript and
every other channel's transcript too.

Same code path also leaked the on-disk transcript: `delete_agent_sessions`
clears `sessions` + `sessions_fts` rows transactionally but never touches
the `<workspace>/sessions/{sid}.jsonl` mirror, so every reset accumulated
orphan files indefinitely.

This change introduces a `ResetScope { Agent, Session(SessionId) }` arg on
`reset_session` / `reboot_session` so the channel layer can ask the kernel
to delete exactly one sid (the one derived from `(channel, chat_id)` via
`SessionId::for_sender_scope`), and adds best-effort JSONL cleanup to all
three session-delete paths (`reset_session`, `reboot_session`,
`clear_agent_history`).

Behaviour summary:
  - Channel `/new` now resets only the per-channel session; sibling
    sessions on other surfaces are untouched. The session is recreated
    eagerly at the same deterministic sid so external `SessionStart` /
    `SessionReset` hooks fire symmetrically and the next inbound message
    lands on a fresh empty row.
  - Per-session reset does NOT clear agent-wide quota state (no grief
    vector for one channel to clear an agent-wide token-quota cap).
  - Cross-agent sids are rejected with `InvalidInput` instead of being
    silently deleted (defence-in-depth on the kernel boundary).
  - `POST /api/agents/{id}/session/reset`, `…/reboot`, the dashboard
    WebSocket `reset` / `reboot` commands, and `clear_agent_history` keep
    the historical agent-wide semantics by passing `ResetScope::Agent`.
  - All three SQL-delete paths now also remove the
    `<workspace>/sessions/{sid}.jsonl` mirror. Failures are logged with
    the orphan path but not propagated — SQL is the source of truth,
    leaving a ghost file is hygiene not privacy.

Concurrency (review fix):
  - All three reset paths (reset / reboot / clear_agent_history) are now
    `async fn` and acquire the same per-agent / per-session message lock
    that `send_message_full` holds. Without this, an in-flight turn's
    later `save_session` UPSERT would silently revive the just-deleted
    row, defeating the entire fix. `reset_one_session` takes
    `agent_msg_locks[agent]` then `session_msg_locks[sid]` (same order
    as `send_message_full`'s lock acquisition — no deadlock can form).

Single source of truth for the scope formula (review fix):
  - The channel-scope formula previously lived inline in four places:
    `kernel/messaging.rs::send_message_full`, `kernel/agent_execution.rs`,
    `kernel/mod.rs::resolve_dispatch_session_id`, and the channel-bridge
    reset helper. Any one drifting from the others silently regresses
    #4868. Extracted to `SessionId::for_sender_scope(agent, channel,
    chat_id)` in `librefang-types` and pinned via a unit test.

Tests:
  - `crates/librefang-kernel/tests/session_reset_scope_test.rs` —
    7 tokio integration tests:
      * per_session_reset_only_deletes_targeted_sid
      * agent_wide_reset_purges_all_sessions_and_jsonl
      * per_session_reset_rejects_cross_agent_sid
      * per_session_reset_on_unmaterialised_sid_is_ok
      * per_session_reboot_skips_summary_save
      * clear_agent_history_purges_jsonl_too
      * compact_with_id_targets_the_requested_session
  - Unit test `for_sender_scope_matches_legacy_inline_formula` in
    `librefang-api/src/channel_bridge.rs` pins the helper's output
    against the pre-refactor inline formula so the four kernel call
    sites can't drift.

Out of scope (deferred per issue's own #3 follow-up): a `librefang sessions
gc` admin tool to reconcile already-orphaned JSONL files on existing
deployments. The new code stops accumulating orphans going forward; the
backfill is a separate environmental cleanup.