fix(tui): persist one-shot session trails for post-mortem debugging

[alpibrupa]

Summary

• `run_once` used `new_session_with_provider`, which opens an in-memory-only trail log (`persist.open_ephemeral`) discarded the moment the process exits. A one-shot run that stops without producing anything — hits its step budget, say — left no record of what it actually did.
• Reproduced live today: a real build task burned its whole 50-step budget on bash exploration with nothing to inspect afterward, not even which commands it ran.
• Switches to `new_session_persistent_with_provider` (already used by graph pipelines), which writes to `.lex/sessions/.db` instead.
• Not a straight swap: `new_session_from_log` always starts a session's in-memory cache at `messages: []`, regardless of what a log under that id already holds — correct for a pipeline node's id (reused deliberately across runs of the same pipeline), but wrong for the fixed `"cli"` id reused across separate one-shot invocations. A second `bin/lex-code "task"` in the same project would find `"cli"`'s log already holding the first run's events, immediately fail the fresh session's `event_count` check (fix(session): per-turn history check is O(1), not O(n) per turn #136), and refuse before ever reaching the model.
• `cli_session_id()` generates a fresh id per invocation instead (timestamp + random suffix), so each run gets its own file with no collision. The one-shot output now names the trail file so a stuck or incomplete run can actually be inspected afterward.
• `repl` is untouched — a live REPL user already watches every step as it happens, which is the case the ephemeral log was originally fine for.

Test plan

• `lex check`/`lex check --strict` on the modified file, plus a full repo sweep
• `lex fmt --check src/`, `lex test` (4/4), `lex doc-sync --check`
• Ran two consecutive one-shot invocations in the same project — both succeeded independently, each with its own trail file, no collision/refusal
• Confirmed post-mortem inspection actually works: `sqlite3` against a completed run's `.db` file shows the full conversation, including the model's own reasoning text

🤖 Generated with Claude Code